Source-linked AI summary
The Machine Learning Landscape of Top Taggers
G. Kasieczka, T. Plehn, A. Butter, K. Cranmer, D. Debnath, B. M. Dillon, M. Fairbairn, D. A. Faroughy, W. Fedorko, C. Gay, L. Gouskos, J. F. Kamenik, P. T. Komiske, S. Leiss, A. Lister, S. Macaluso, E. M. Metodiev, L. Moore, B. Nachman, K. Nordstrom, J. Pearkes, H. Qu, Y. Rath, M. Rieger, D. Shih, J. M. Thompson, S. Varma
TL;DR
Top tagging provides an established benchmark for comparing modern machine-learning approaches on low-level jet information. The paper compares image-, four-vector-, and theory-motivated taggers and finds comparable performance across these architectures, while identifying experimental and theoretical limitations and open practical questions.
Problem
The study asks how modern neural networks can classify jets from low-level calorimeter information and which approaches use the available information most effectively.
Method
The paper compares image-based, four-vector-based, and theory- or relativistic-kinematics-motivated deep-learning taggers for hadronic top classification.
Results
The three tagger classes provide competitive performance with comparable results, indicating no single golden deep-network architecture.
Takeaways & Limitations
Simple tagging performance cannot select the preferred architecture, so future evaluation should address versatility, stability, calibration, and uncertainty estimates.
Takeaways & Limitations
The dataset omits particle-level and tracking information, and particle-flow TopoDNN results are not directly comparable to the cited ATLAS results.
Abstract
from arXiv · showhide
Based on the established task of identifying boosted, hadronically decaying top quarks, we compare a wide range of modern machine learning approaches. Unlike most established methods they rely on low-level input, for instance calorimeter output. While their network architectures are vastly different, their performance is comparatively similar. In general, we find that these new approaches are extremely powerful and great fun.
1 Introduction
Top tagging is an established, experimentally relevant benchmark, and deep learning extends it from engineered observables toward low-level LHC data. The study compares modern neural-network approaches and finds substantial performance improvement, while raising questions about training data, benchmarking, calibration, and systematics.
- Top tagging identifies boosted hadronic top quarks and is a benchmark candidate because it is experimentally and theoretically well defined.
- Deep learning enables classification of low-level LHC data without constructing high-level observables.
- The study compares calorimeter-image, particle four-momentum, recurrent, and theory-motivated taggers.
- Modern machine-learning tools consistently improve performance over earlier multivariate top-tagging methods.
- The resulting performance gains motivate further questions about training data, benchmarking, calibration, and systematic uncertainties.
2 Data set
The dataset provides matched signal and background jets simulated at 14 TeV, represented primarily through calorimeter-derived information. It uses fixed training, validation, and test samples, while omitting particle charge and tracking information.
- The samples contain top signal and mixed quark-gluon background jets generated with Pythia8 at 14 TeV and passed through Delphes detector simulation.
- Signal jets require a matched parton-level top and all top decay partons within ΔR = 0.8 of the jet axis.
- The format excludes particle charge and additional tracking information, so its performance is not the final word for LHC top tagging.
- The public dataset contains 1 million signal and 1 million background jets, split into training, validation, and testing samples.
- Results are evaluated on the test sample, while per-jet classification outputs support comparisons of which jets each algorithm classifies correctly.
3.1 Imaged-based taggers
Image-based taggers convert calorimeter energy deposits into jet images and apply convolutional networks. The compared CNN approaches use different image resolutions and architectures, with preprocessing designed to align jet structure.
- Calorimeter energy deposition inside a fat jet is represented as a sparsely filled image for convolutional-network analysis.
- A fat jet with R = 0.8 can be represented by 40 × 40 pixels at the stated calorimeter resolution.
- Standard CNN preprocessing centers and rotates jets using the pT-weighted centroid and principal axis, then flips them before pixelation.
- The CNN uses grayscale images whose pixel intensity is pT, normalized to unit total intensity.
- The study’s CNN extends the DeepTop architecture with more feature maps and hidden units.
- ResNeXt uses 64 × 64 jet images and a modified 50-layer ResNeXt architecture with reduced convolutional channels and dropout.
3.2 4-Vector-based taggers
4-vector-based taggers use constituent-level jet information in several architectures, including dense, N-subjettiness, tree-based, convolutional, and graph-based networks. These approaches encode jet substructure through prongs, momenta, constituent relations, or learned tree and point-cloud representations.
- Particle-flow inputs address the difficulty of combining calorimeter information with tracking or particle identification at mismatched resolutions.
- TopoDNN: TopoDNN uses up to 200 pT-sorted constituents represented by (pT, η, φ), with physics-motivated centering, rotation, and momentum scaling.The study finds that 30 constituents saturate the network performance, while the ATLAS comparison uses different inputs and is not direct.
- TopoDNN: A proper three-dimensional rotation aligns the second-highest-pT constituent with the negative y-axis, preserving jet mass while potentially distorting N-subjettiness.
- Multi-Body N-Subjettiness: The multi-body phase-space tagger feeds N-subjettiness variables into a dense neural network, using kT axes with N = 6 or 8.The variables are intended to span an m-body phase space and combine prong-counting with momentum information in an infrared-safe input.
- TreeNN and TreeNiN: TreeNN embeds reclustered jet trees into fixed-size vectors, while TreeNiN adds fully connected layers at each binary-tree node.The tree topology is determined by the clustering algorithm; this study uses kT clustering.
- Particle-Based Networks: P-CNN processes up to 100 pT-ranked constituents with seven transformed features, whereas ParticleNet applies dynamic graph convolutions to unordered constituent point clouds.ParticleNet updates nearest-neighbor edges between stages and applies global pooling followed by fully connected layers.
3.3 Theory-inspired taggers
Theory-inspired taggers incorporate physics-motivated structure into representations of jet constituents, aiming to improve efficiency, interpretability, or stability. The approaches include Lorentz-based feature extraction, graph-inspired transformations, latent-theme modeling, and IRC-safe observable bases.
- Motivation: Physics-informed features can reduce training time and reveal which information a network uses.The motivation is to encode subjet-analysis knowledge directly into the network.
- Physics motivation: Top and massless QCD jets differ at four-vector level through two mass drops after soft- and collinear-based recombination.Providing this structure explicitly may be more efficient than having taggers learn Lorentz boosts and the Minkowski metric.
- Infrared safety: Jet taggers should remain stable under additional soft or collinear splittings, but particle-level outputs are not clearly infrared safe.This uncertainty affects comparisons with QCD predictions and the training, benchmarking, and Monte Carlo extraction of tagging efficiencies.
- Lorentz Boost Network: Lorentz-based networks reconstruct parent particles from constituent four-vectors and use their properties in appropriate rest frames.The LBN combines inputs with trainable positive coefficients before extracting physics-motivated features and jointly training the downstream classifier.
- Lorentz Layer: The LoLa network combines input four-vectors and transforms them into measurement-motivated objects using sums or minima of Minkowski distances.Its setup uses N = 60 input vectors and M = 90 resulting vectors before fully connected layers.
- Latent themes and observables: LDA learns two signal- and background-oriented themes from co-occurrence patterns in high-level jet-substructure features, while EFPs provide an IRC-safe observable basis for linear models.The EFP basis is organized by graph edge count, with exactly 1000 EFPs for d ≤7.
4 Comparison
Across the evaluated taggers, performance is broadly similar and strong, with several approaches reaching approximately 0.98 AUC and exceeding 1/300 signal-to-background improvement. Ensemble analyses and the GoaT combination show additional gains, while the study’s conclusions are bounded by its simulated, calorimeter-focused setup and narrow jet pT range.
- 93% typical accuracy and similar ROC shapes indicate that the quoted differences are unlikely to define a clear experimental preference.The ROC comparison uses the same test sample and includes a simple N-subjettiness-plus-mass tagger baseline.
- Most approaches achieve an AUC of approximately 0.98, with ParticleNet, ResNeXt, TreeNiN, and PFN leading the reported metrics.These approaches also reach the highest accuracy and background rejection at fixed signal efficiency.
- 5 ... 15% ensemble performance improvement is observed depending on the algorithm, with background rejection ranging from around 1/600 to better than 1/1000.The ensemble setup trains nine models per tagger and combines six at a time; the leading ensemble approaches are ParticleNet, ResNeXt, TreeNiN, and PFN.
- ϵS/ϵB > 300 is achieved by the leading ensemble taggers, vastly exceeding current ATLAS and CMS top-tagging performance.This improvement is reported in terms of the signal-to-background ratio.
- More than 10% higher background rejection is obtained by the GoaT combination than by the best individual tagger, estimating remaining information in classifier correlations.GoaT is presented as a benchmark of unused correlated information rather than a prospective analysis tool.
- The conclusions are limited by a narrow and specific jet pT slice, simple detector simulation without underlying event or pile-up, calorimeter-only inputs, and unaddressed computational costs.Systematic uncertainties when applying classifiers trained on simulation to data are also identified as an important open issue.
5 Conclusion
Top tagging provides a well-defined LHC benchmark for comparing deep-learning architectures and data representations. Image-based, 4-vector-based, and theory-informed approaches achieve comparable performance, so practical criteria beyond tagging accuracy remain important.
- Top tagging is a prime benchmark for modern machine-learning classification tasks relevant to the LHC.The task distinguishes hadronically decaying top quarks from light-quark or gluon jets.
- Three deep-learning classes—image-based, 4-vector-based, and theory-informed taggers—provide competitive performance.
- Comparable performance across architectures means simple tagging accuracy cannot identify the best network design.
- Future comparisons should examine versatility, stability, calibration, and uncertainty estimates in the specific LHC environment.