Source-linked AI summary
Playing Tag with ANN: Boosted Top Identification with Pattern Recognition
Leandro G. Almeida, Mihailo Backovic, Mathieu Cliche, Seung J. Lee, Maxim Perelstein
TL;DR
Boosted top tagging is difficult because collimated top decays must be separated from common QCD jets. The paper treats calorimeter energy deposits as images and trains an ANN tagger, which outperforms existing taggers in Monte Carlo studies while remaining subject to realistic-environment limitations.
Problem
Boosted top tagging is needed for LHC searches because heavy-particle decays produce highly collimated top-quark decay products that are difficult to reconstruct.
Method
The paper trains an Artificial Neural Network on labeled top and non-top jet samples using calorimeter energy information represented as digital images.
Results
60% tagging efficiency is achieved with about 4% mis-tag rate for jets with pT ∈[1.1, 1.2] TeV, about a factor of 2 lower than the best comparison tagger.
Takeaways & Limitations
The ANN approach shows promise for improving boosted top-tagging performance and uses physical jet features that generalize across showering models.
Takeaways & Limitations
The study is an initial analysis that does not fully capture the complexity of a realistic experimental environment.
Abstract
from arXiv · showhide
Many searches for physics beyond the Standard Model at the Large Hadron Collider (LHC) rely on top tagging algorithms, which discriminate between boosted hadronic top quarks and the much more common jets initiated by light quarks and gluons. We note that the hadronic calorimeter (HCAL) effectively takes a "digital image" of each jet, with pixel intensities given by energy deposits in individual HCAL cells. Viewed in this way, top tagging becomes a canonical pattern recognition problem. With this motivation, we present a novel top tagging algorithm based on an Artificial Neural Network (ANN), one of the most popular approaches to pattern recognition. The ANN is trained on a large sample of boosted tops and light quark/gluon jets, and is then applied to independent test samples. The ANN tagger demonstrated excellent performance in a Monte Carlo study: for example, for jets with p_T in the 1100-1200 GeV range, 60% top-tag efficiency can be achieved with a 4% mis-tag rate. We discuss the physical features of the jets identified by the ANN tagger as the most important for classification, as well as correlations between the ANN tagger and some of the familiar top-tagging observables and algorithms.
1 Introduction
The paper frames boosted top identification as an image-recognition problem and introduces an ANN tagger trained to distinguish top jets from non-top jets.
- Motivation: Boosted top decays produce highly collimated quarks and distinctive hadronic-activity patterns that challenge classical reconstruction.For pT ∼1 TeV, the three quarks occupy a cone of roughly R ∼0.4.
- Related approaches: Jet-substructure methods use energy-distribution observables and clustering histories to distinguish heavy-particle jets from light-parton QCD jets.
- Image-recognition viewpoint: The HCAL records each jet as a digital image whose pixels are calorimeter cells and whose intensities are deposited energies.
- ANN approach: The ANN classifies each jet as top or non-top with a nonlinear scoring function whose weights are learned from labeled training samples.
- Reported result: On independent samples, the ANN significantly outperforms several popular tagging algorithms across a wide range of jet pT.
2 Event Generation and Pre-Processing
The study generates Monte Carlo jet samples and preprocesses each jet into a centered, oriented, normalized calorimeter image for ANN input.
- Event generation: Benchmark events are generated with MadGraph 5 at leading order and showered with Pythia 6, with separate Pythia 8 samples used to study showering effects.
- Event selection: A jet-mass selection retains most top jets while rejecting most QCD jets before the ANN discriminates the remaining background.
- Coordinate preprocessing: The jet center is defined from energy-weighted particle coordinates, then each particle is shifted so the jet is centered at the origin.
- Coordinate preprocessing: The coordinate system is rotated so every jet’s principal axis points in the same +η direction.
- Coordinate preprocessing: Centering and rotation remove jet position and orientation information, allowing classification to focus on physical differences between top and QCD jets.
- Image construction: The transformed window is divided into 30 × 30 cells, and each cell stores energy normalized by the jet’s total energy.Approximately 98% of assigned particles lie in the window, whose cell size is about 0.1 × 0.1.
3 ANN Tagger
The ANN uses normalized calorimeter images as input, learns through backpropagation, and controls over-fitting with sufficient training data and an ensemble of networks.
- ANN architecture: The network has 900 input nodes, two 100-node hidden layers, and one output node interpreted as the probability of a boosted top decay.
- ANN architecture: The 900 inputs are the normalized 30 × 30 cell energies rearranged into a single vector, while weights and biases are learned during training.
- Training: Training assigns target outputs of 1 to top jets and 0 to QCD jets, then adjusts weights to reduce logarithmic loss.
- Training: Backpropagation with gradient descent minimizes the training loss by propagating output errors backward through the network.
- Over-fitting control: Performance improves with training-set size until Ntr ≈40000, after which convergence indicates minimal over-fitting.This corresponds to 20000 top images and 20000 dijet images.
- Ensembling: An ensemble averages the outputs of B = 10 neural networks, reweighting difficult examples so later networks focus on misclassified jets.
4 Results
The ANN separates top and QCD jets using calorimeter-image patterns, outperforming existing taggers across representative transverse-momentum ranges and remaining effective under alternative showering models.
- ANN performance: About 50,000 top and QCD jets per pT bin trained the ANN ensemble, which was tested on independent samples of about 15,000 jets per class.The three bins were 500–600, 800–900, and 1100–1200 GeV.
- ANN performance: 60% tagging efficiency with about 4% mis-tag rate was achieved for jets with pT ∈[1.1, 1.2] TeV.This mis-tag rate was about a factor of 2 lower than the best existing tagger in the comparison pool.
- Comparison with existing taggers: The ANN outperformed the d12, top template, and N-subjettiness taggers, achieving lower mis-tag rates at the same tagging efficiency.Efficiency-versus-mis-tag curves were compared across three representative pT ranges.
- Jet features: High-scoring top jets showed a well-defined three-prong structure, whereas low-scoring examples had nearly collinear quarks or one much softer decay product.The corresponding figure compares energy-deposit patterns for the highest- and lowest-scoring top jets in the 800–900 GeV range.
- Jet features: High-scoring QCD jets contained well-defined isolated subjets and were therefore more likely to be misidentified as tops.The figure contrasts lowest- and highest-scoring QCD jets in the 800–900 GeV range.
- Correlations and robustness: ANN scores had significant but imperfect positive correlations with other taggers, with coefficients ranging from about 0.3 to 0.7.The ANN and N-subjettiness rankings were correlated in both top and light-jet samples, while the ANN still achieved better efficiency-versus-mis-tag performance.
- Mass dependence: For jets below 130 GeV mass, positive-tag probability dropped rapidly for both top and QCD samples, while above 210 GeV it was roughly mass-independent.The ANN rejected low-invariant-mass jets, a behavior noted as potentially useful for reducing pile-up effects.
- Correlations and robustness: The ANN remained effective on Pythia 8 samples despite training on Pythia 6, suggesting its classification features were physical rather than shower-model artifacts.Generator dependence affected all compared taggers by a similar amount.
5 Discussion
The ANN top tagger significantly outperforms traditional taggers in the Monte Carlo study, but its realistic experimental applicability remains limited by simulation scope, missing detector effects, and a simple architecture.
- Discussion: The ANN tagger significantly outperforms traditional taggers on the Monte Carlo datasets studied.The authors suggest it may serve as a benchmark for other boosted top-tagging observables.
- Limitations: The study does not yet fully capture the complexity of realistic experimental conditions.It is presented as a first study of the approach.
- Limitations: The simulations include only leading-order t¯t and dijet processes, leaving possible accidental substructure from higher-multiplicity QCD events unmodeled.The authors state that subleading processes may have an outsized effect on tagger performance.
- Limitations: Pile-up is absent from the simulations, and its interaction with the ANN tagger remains unexplored.Methods for reducing pile-up effects exist, but their combination with this tagger requires study.
- Limitations: Possible Monte Carlo biases in ANN training must be addressed before applying the method to real data.A preliminary study finds the determining features are not strongly MC-dependent, but more extensive validation with real LHC control samples is needed.
- Future directions: The tagger uses only HCAL information and a relatively simple neural-network architecture, while b-tagging and advanced network techniques may improve performance.The paper mentions convolutional networks and unsupervised pre-training as possible improvements.
- Future directions: The pattern-recognition approach may also apply to boosted W and h jets, although improvements over traditional taggers remain to be tested.The paper focuses exclusively on tops and identifies these extensions as an interesting direction.
A A Brief Description of Existing Top Taggers
The comparison uses three existing top-tagging strategies: template matching, N-subjettiness prong observables, and clustering-history information, with specified implementation choices.
- Overview: The comparison includes three existing methods, each exploiting a different approach to boosted top tagging.They are used as representative algorithms against the ANN tagger.
- Template Overlap Method: Template Overlap matches a fat jet’s energy distribution to libraries of kinematically allowed parton-level heavy-particle decays.The overlap function quantifies the quality of the match.
- Template Overlap Method: The template libraries use fixed total transverse momentum, a 40-step η–φ scan, and an angular region of R = 1.0 around the fat-jet axis.These choices are made for simplicity and processing speed.
- N-subjettiness: N-subjettiness calculates τN moments by minimizing pT-weighted distances between calorimeter deposits and trial axes dividing the jet into N regions.For top tagging, the comparison uses the ratio τ3/τ2 with angular exponent β = 1.
- ATLAS top tagger: The ATLAS top tagger uses d12, the kT norm at the clustering step where two subjets merge into one final jet.The comparison omits the usual lower trimmed-mass cut because its samples already satisfy a jet-mass restriction.