Source-linked AI summary
Deep-learning Top Taggers or The End of QCD?
Gregor Kasieczka, Tilman Plehn, Michael Russell, Torben Schell
TL;DR
The paper examines whether convolutional neural networks can match established QCD-based top taggers for identifying top quarks in fat jets. It optimizes a DeepTop network on jet images from Monte Carlo simulations and compares it with multivariate QCD-based taggers. The approaches achieve comparable performance, supporting convolutional networks as a promising approach for multivariate hypothesis-based top tagging.
Problem
The study asks how machine-learning top taggers compare with established, well-performing QCD-based top taggers in a controlled fat-jet setting.
Method
The authors represent calorimeter energy deposits as jet images, train convolutional neural networks for top tagging, optimize the architecture, and compare the resulting networks with QCD-based taggers.
Results
The neural-network and QCD-based approaches have comparable performance.
Takeaways & Limitations
Convolutional networks provide a promising new approach for multivariate hypothesis-based top tagging.
Takeaways & Limitations
The neural network’s performance begins to degrade when pixels below approximately 5 GeV transverse energy are excluded, indicating sensitivity to low-pT information.
Abstract
from arXiv · showhide
Machine learning based on convolutional neural networks can be used to study jet images from the LHC. Top tagging in fat jets offers a well-defined framework to establish our DeepTop approach and compare its performance to QCD-based top taggers. We first optimize a network architecture to identify top quarks in Monte Carlo simulations of the Standard Model production channel. Using standard fat jets we then compare its performance to a multivariate QCD-based top tagger. We find that both approaches lead to comparable performance, establishing convolutional networks as a promising new approach for multivariate hypothesis-based top tagging.
1 Introduction
The paper asks whether convolutional neural networks can analyze fat-jet substructure as effectively as established QCD-based top taggers. It represents calorimeter deposits as jet images and combines convolutional feature extraction with neural-network classification.
- Motivation: The study compares machine-learning top tagging with state-of-the-art QCD-based taggers in a well-defined fat-jet environment.This comparison addresses whether QCD-based taggers remain useful or may eventually be replaced by pattern recognition.
- Jet images: Convolutional neural networks treat the rapidity–azimuth plane as a sparsely filled image whose pixel intensities encode calorimeter energy deposits.Signal and background jet images are preprocessed and supplied to the network for learning.
- Classification: The network’s final layer converts learned image features into a signal-or-background probability.Performance is evaluated with receiver operating characteristic curves, analogously to multivariate top-tagger analyses.
- QCD-based taggers: QCD-based comparisons use boosted decision trees, whose sequentially reweighted tree collection classifies events through the aggregate vote of its classifiers.These BDTs are part of the standard LHC toolbox and are used for the QCD-based taggers.
- Network architecture: Convolution layers apply local filters to neighboring pixels, producing feature maps that feed later convolution layers and a final one-dimensional DNN.Pooling between convolutional blocks reduces feature-map size, while the DNN performs the final classification.
- Training: Training minimizes a mean squared error using gradient-based parameter updates, with learning rate, momentum, and optionally Nesterov-style gradient evaluation controlling optimization.The stated loss uses predicted binary labels and true labels; the authors note that it is sufficient for their purpose rather than optimal for learning performance or probabilistic information.
2 Machine learning setup
The setup represents calorimeter energy deposits as pre-processed jet images and trains an optimized ConvNet/DNN for fat-jet top tagging. The study uses Monte Carlo samples, defined fat-jet selections, and layer-level diagnostics to examine both performance and learned patterns.
- The analysis aims to test whether machine-learning top tagging offers an advantage over established QCD-based taggers and to understand the ConvNet’s learning pattern.
- The study simulates 14 TeV hadronic t-tbar and QCD dijet samples, using fast detector simulation, R=1.5 anti-kT fat jets, and central jets with |η|<1.0.
- Signal fat jets are restricted to pT,fat=350 ... 450 GeV and must match a Monte Carlo truth top within ΔR<1.2 so decay products are captured.
- Jet images and pre-processing: The full setup averages 10,000 signal and background images, placing the leading subjet centrally, the second at 12 o’clock, and a smeared third signal subjet on the right.
- Network architecture: The architecture scan splits 600,000 signal and 600,000 background images into training, validation/optimization, and final-test samples.
- Network architecture: The ConvNet uses repeated zero-padding, convolution, and activation layers followed by pooling, then feeds learned feature maps into dense DNN layers.The setup uses ReLU activations and Glorot-uniform weight initialization, with the default network selected for validation performance and hyperparameter stability.
3 Performance test
The performance test compares DeepTop with QCD-based multivariate top taggers on simulated fat-jet samples. The approaches achieve close performance, while learned observables and preprocessing effects clarify what drives the comparison.
- 3 Performance test: The benchmark separates boosted top decays inside fat jets from QCD fat jets in the transverse-momentum range pT,fat = 350 ... 450 GeV.The Monte Carlo training and testing sample replaces actual data for this performance comparison.
- 3.1 QCD-based taggers: The QCD shower-based taggers do not fully use available calorimeter information, motivating the addition of N-subjettiness variables.The τ3/τ2 observable is particularly useful with QCD taggers in a multivariate setup.
- 3.1 QCD-based taggers: The QCD-based comparison uses SoftDrop with N-subjettiness, with additional HEPTopTagger information included in the MotherOfTaggers BDT.SoftDrop uses β = 1 and zcut = 0.2; HEPTopTagger targets three-prong top structure and reconstructed top mass.
- 3.2 Comparison: Minimal preprocessing performs slightly better than full preprocessing, because rotation and cropping introduce a small information loss.The preprocessing difference is comparable in size to the difference between tagging approaches.
- 3.2 Comparison: The two QCD-based BDT analyses and both DeepTop neural-network setups are close in performance.Adding HEPTopTagger information slightly improves SoftDrop+N-subjettiness in the studied transverse-momentum range.
- 3.2 Comparison: DeepTop and MotherOfTaggers learn essentially the same kinematic structures in mfat and τ3/τ2 distributions.MotherOfTaggers signal curves are more peaked because these observables are direct BDT inputs, whereas DeepTop derives them.
- 3.2 Comparison: The neural network performance begins to degrade when retaining only pixels with transverse energy deposition ET > 5 GeV.This cutoff tests the extent to which the network relies on low-pT effects.
4 Conclusions
The conclusions establish convolutional networks as a viable approach for fat-jet top tagging and compare them directly with QCD-based multivariate taggers. DeepTop and the QCD-based approach show comparable performance, while the network remains interpretable and stable in the tested studies.
- 4 Conclusions: Fat jets containing boosted hadronic top decays provide a well-defined setting for comparing machine learning with QCD-based top taggers.The study treats top tagging as testing QCD versus top-decay hypotheses rather than reconstructing an actual top decay.
- 4 Conclusions: Network architecture changes have only a small impact on top-tagging performance.Independent event samples are used to optimize, train, and test the network.
- 4 Conclusions: DeepTop and the MotherOfTaggers QCD-based tagger have comparable performance in the benchmark.This comparison leaves freedom among reconstruction, multivariate hypothesis testing, and data-based machine-learning strategies.
- 4 Conclusions: The neural network is not dominated by low-pT calorimeter entries and is extraordinarily stable under changes to the jet energy scale.The network also learns kinematic observables in the fat jet, including structures examined through the comparison.
A What the machine learns
DeepTop and the QCD-based BDT extract broadly consistent jet-substructure information, although their responses differ in selected mass and transverse-momentum observables.
- Mass and substructure observables: SoftDrop mass identifies the correct top mass in fewer than half of correctly identified signal events, whereas fat jet mass reproduces it.SoftDrop remains useful because its background distribution peaks around msd ≈20 GeV, enabling a hypothesis test for massive decays.
- Mass and substructure observables: The DNN and BDT produce consistent results for the mass, optimalR, and N-subjettiness observables.The BDT performs slightly better for clear signal events, with a more pronounced background output.
- Mass and substructure observables: The HEPTopTagger W-to-top mass ratio frec has little significance in the studied transverse-momentum range.The optimalR variable ∆Ropt and τ3/τ2 show consistent results between the DNN and BDT.
- Transverse-momentum response: Classifier-selected transverse-momentum distributions reveal a tagger-response bias, larger for the DNN than for the BDT.The un-groomed fat jet reproduces the Monte Carlo range pT,fat = 350 ... 450 GeV, while SoftDrop alters the constituent spectrum.
- Robustness checks: At a signal efficiency of 30 %, background rejection is evaluated across activation thresholds and image rescalings.These studies probe how the network response changes under thresholding and jet-energy-scale variations.
B Detector effects
The network is relatively tolerant of removing very soft calorimeter activity, but its tagging performance changes under global jet-energy-scale rescaling.
- Activation thresholds: Removing pixel activity below 3 GeV only slightly degrades network performance.Above 3 GeV, background rejection decreases approximately linearly as the activation threshold increases.
- Jet-energy-scale effects: A ±25% global jet-energy-scale rescaling changes tagging performance by approximately −10% when reduced and −5% when increased.The evaluation rescales ET pixel intensities globally to assess systematic sensitivity.
- Jet-energy-scale effects: Random image rescaling during hardened training preserves similar performance while further reducing sensitivity to global jet-energy-scale changes.The rescaling uses a Gaussian distribution with mean 1.0 and width 0.1, with new random values each epoch.
- Experimental implications: The network’s resilience and ability to be hardened are encouraging for experimental use, although non-uniform image distortions remain to be considered.Understanding and mitigating systematic uncertainties is identified as critical for experimental applications.