Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

The paper benchmarks a broad spectrum of deep learning taggers for identifying boosted hadronically decaying top quarks using low-level calorimeter input. It categorizes taggers into image-based, 4-vector-based, and theory-inspired approaches, showing broadly similar performance with ROC-AUC around 0.98 and strongest results from ParticleNet and ResNeXt. It also investigates ensembles and a meta-tagger (GoaT), revealing incremental gains and underscoring remaining challenges in calibration, uncertainty treatment, IR-safety, and detector effects. Overall, it demonstrates that deep networks operating on raw jet information provide competitive, versatile tools for LHC jet tagging, while stressing practical considerations for real-world deployment and further robustness studies.

Abstract

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.

The Machine Learning Landscape of Top Taggers

TL;DR

The paper benchmarks a broad spectrum of deep learning taggers for identifying boosted hadronically decaying top quarks using low-level calorimeter input. It categorizes taggers into image-based, 4-vector-based, and theory-inspired approaches, showing broadly similar performance with ROC-AUC around 0.98 and strongest results from ParticleNet and ResNeXt. It also investigates ensembles and a meta-tagger (GoaT), revealing incremental gains and underscoring remaining challenges in calibration, uncertainty treatment, IR-safety, and detector effects. Overall, it demonstrates that deep networks operating on raw jet information provide competitive, versatile tools for LHC jet tagging, while stressing practical considerations for real-world deployment and further robustness studies.

Abstract

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.

Paper Structure

This paper contains 21 sections, 8 equations, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Data set
  3. Taggers
  4. Imaged-based taggers
  5. CNN
  6. ResNeXt
  7. 4-Vector-based taggers
  8. TopoDNN
  9. Multi-Body N-Subjettiness
  10. TreeNiN
  11. P-CNN
  12. ParticleNet
  13. Theory-inspired taggers
  14. Lorentz Boost Network
  15. Lorentz Layer
  16. ...and 6 more sections

Figures (6)

  • Figure 1: Left: typical single jet image in the rapidity vs azimuthal angle plane for the top signal after pre-processing. Center and right: signal and background images averaged over 10,000 individual images.
  • Figure 2: Architecture of the CNN top tagger. Figure from Ref. Macaluso:2018tck.
  • Figure 3: Number of constituents (left) and mean of the transverse momentum (right) of the ranked constituents of a typical top jet. We show calorimeter entries as well as particle flow constituents after Delphes.
  • Figure 4: Visualization of the trained top tagging EFN. Each contour corresponds to a filter, which represents the learned local latent space. The smaller filters probe the core of the jet and larger filters the periphery. Figure from Ref. Komiske:2018cqr.
  • Figure 5: ROC curves for all algorithms evaluated on the same test sample, shown as the AUC ensemble median of multiple trainings. More precise numbers as well as uncertainty bands given by the ensemble analysis are given in Tab. \ref{['tab:overview']}.
  • ...and 1 more figures