Pith. sign in

REVIEW 4 major objections 5 minor 36 references

Vision Transformers for End-to-End Quark-Gluon Jet Classification from Calorimeter Images

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that vision transformer hybrids, particularly ViT+MaxViT and ViT+ConvNeXt, outperform standard CNNs for quark-gluon jet classification from multi-channel calorimeter images, and it establishes a structured benchmark…

desk verdict Useful benchmark, but the ViT-superiority claim is contradicted by the paper's own table and confounded by unequal augmentation. read the letter →

arxiv 2506.14934 v1 pith:K2Z27LYA submitted 2025-06-17 cs.CV

Md Abrar Jahin , Shahriar Soudeep , Arian Rahman Aditta , M. F. Mridha , Nafiz Fahad , Md. Jakir Hossen This is my paper
classification cs.CV MSC 68T0768T45
keywords visiontransformersquark-gluonjetsjettaggingcalorimeterimagesCMSOpenDataend-to-endlearninghybridneuralnetworkstransfer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that Vision Transformer architectures, which model global context across an image, classify quark-gluon jets from raw calorimeter images better than convolutional neural networks. The authors build three-channel jet images from simulated 2012 CMS Open Data, encoding electromagnetic calorimeter deposits, hadronic calorimeter deposits, and reconstructed tracks in a single grid. They benchmark standalone CNNs, transformers, and hybrid ViT-CNN models; the top performers, ViT+MaxViT and ViT+ConvNeXt, beat every pure CNN baseline on F1-score, ROC-AUC, and accuracy. The paper also contributes a structured dataset of 933,206 images and a training pipeline for future work. A sympathetic reading of the claim is that attention-based models merit a place in the standard toolkit for end-to-end jet tagging on calorimeter images.

What carries the argument

The carrying object is the multi-channel jet image: a 125×125 pixel grid in the (η, φ) plane with three channels (ECAL, HCAL, tracks), each pixel weighted by transverse momentum. The decisive mechanism in the hybrids is feature concatenation: a ViT-Base/16 backbone pre-trained on a large generic image corpus produces a global class-token embedding, which is concatenated with a convolutional backbone's pooled features and passed through a two-layer MLP head. The paper argues that this design supplies both global attention and convolutional locality, and that the attention component is what lets the hybrids break past CNN performance.

What would settle it

Retrain the CNN baselines with the same Mixup augmentation and preprocessing used for the transformers, then evaluate all models on a stratified sample of the full 933,206-image dataset; if the CNNs match or surpass the ViT+MaxViT and ViT+ConvNeXt hybrids, the claimed attention advantage would be shown to stem from unequal training conditions rather than from the architecture.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that global attention improves quark-gluon discrimination on detector images. In a benchmark of over a dozen architectures, ViT+MaxViT achieves the best F1 score (72.02%) and ROC-AUC (76.65%), while ViT+ConvNeXt achieves the best accuracy (70.57%), both exceeding every pure CNN baseline and the best standalone transformer on all three headline metrics. The authors attribute the margin to attention's ability to capture long-range spatial correlations in jet substructure, patterns that convolutional receptive fields are too local to resolve.

Load-bearing premise

The comparison assumes all models are trained under equivalent conditions and evaluated on a representative sample, but the paper applies Mixup only to ViT and Swin models and reports results on an unspecified 7,000-image subset of the full dataset.

Editorial extensions

If this is right

  • Transformer-based and hybrid models should be adopted as baselines in future image-based jet-tagging studies, alongside CNNs.
  • The released 933,206-image dataset gives researchers a shared benchmark for testing attention mechanisms under realistic pileup conditions.
  • Attention models' higher ROC-AUC at fixed false-positive rates implies they could improve signal sensitivity in physics searches that rely on quark-gluon discrimination.
  • The hybrid feature-concatenation recipe (ViT plus a convolutional backbone) appears to be a broadly effective way to combine global and local image reasoning for detector data.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the attention advantage is real, it should grow with pileup, since off-axis background that confuses local filters is exactly the long-range structure attention can track; this is a testable prediction the paper does not make.
  • A fair re-run that applies Mixup to the CNN baselines as well could narrow or close the gap; the paper's asymmetric augmentation is a confound it does not address.
  • The same multi-channel image pipeline could be transferred to other jet-tagging tasks, such as W/Z or top tagging, and to current higher-luminosity detector geometries.
  • Swapping the generic-image pretrained backbones for self-supervised or physics-specific pretraining might improve data efficiency, a direction the paper leaves open.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. This paper benchmarks Vision Transformer (ViT), CNN, and ViT-CNN hybrid architectures for quark-gluon jet classification using simulated 2012 CMS Open Data. The authors construct 125×125 multi-channel jet images from ECAL, HCAL, and track information, train models with a common preprocessing and training pipeline, and report that ViT+MaxViT and ViT+ConvNeXt hybrids achieve the highest F1 score, ROC-AUC, and accuracy. The paper also presents a sensitivity analysis over dataset size, model variant, batch size, learning rate, optimizer, weight decay, and dropout.

Significance. If the central claim were fully supported, the paper would provide a useful public benchmark for transformer-based jet image classification with realistic detector conditions. The strengths are the use of public CMS Open Data, the construction of multi-channel jet images, the release of code on GitHub, and the breadth of architectures compared. However, the headline claim that 'ViT-based models consistently outperform' CNNs is contradicted by the paper's own Table 1 for standalone ViT and Swin models, and the comparison is confounded by unequal augmentation. These issues are central rather than cosmetic, so the significance of the work as stated is not established.

major comments (4)
  1. [Abstract; §4.1, Table 1] The abstract claims that 'ViT-based models ... consistently outperform established CNN baselines in F1-score, ROC-AUC, and accuracy,' but Table 1 shows standalone ViT (ROC-AUC 69.28±0.04) and Swin (69.28±0.06) substantially below RegNetY (75.89±0.02) and several other CNN baselines. Only the two hybrid ensembles ViT+MaxViT and ViT+ConvNeXt occupy the top ranks. The claim should be narrowed to hybrids, or additional evidence must be provided for standalone transformer advantage.
  2. [§3.2, Eqs. (12)–(13); Table 1] Mixup augmentation is applied only to ViT and Swin models, while CNN baselines use only ImageNet normalization (Eq. 11). Since Mixup is a strong regularizer that typically improves generalization, the observed performance gap between the top hybrid models and pure CNNs could be entirely due to unequal training procedures rather than to attention-based architectures. A controlled experiment with identical augmentation and regularization across all models is required before any architecture-level conclusion can be drawn.
  3. [Table 1 caption; §3.1] Table 1 reports results on '7k samples' from the 933,206-image dataset, but the manuscript does not specify how these samples were selected, whether they are drawn from the training or validation portion, or how the stated 80:20 train/validation split from §3.1 relates to this subset. Without this sampling protocol, the representativeness of the ranking is unclear and the reader cannot assess the stability of the reported ordering.
  4. [§4.2, Figure 2] The sensitivity analysis is not tied to a clearly identified model, and it contains text that is unrelated to jet classification: the first paragraph of §4.2 refers to 'complex medical imaging tasks' and to ViT-Huge as advantageous for such tasks, which is not part of this study. In addition, Figure 2g mentions unfreezing the ViT backbone at epoch 6 and the MaxViT backbone at epoch 10, whereas §3.4 specifies unfreezing at epochs 5 and 8; the optimizer Lion and the ViT-Huge variant are not described in the experimental setup. These inconsistencies undermine the reliability of the ablation claims.
minor comments (5)
  1. [§3.4] The cosine annealing schedule is described as having 'a maximum of 50 iterations,' which appears to be a typo for epochs; please clarify.
  2. [Eq. (5)] In the outlier clipping step, the notation '500×σ_k' is confusing because the data have already been standardized to unit variance per channel, making the threshold numerically equal to 500; please state the intended threshold directly.
  3. [§3.3] The text describing ViT-Base/16 contains rendering artifacts ('224ÃŮ224' and '16ÃŮ16') that should be corrected to '224×224' and '16×16'.
  4. [References] Reference [19] appears to be a self-citation that is not directly relevant to the Lorentz-equivariant network discussion; please verify whether it supports the statement or remove it.
  5. [Figure 2] The x-axis of panel (d) is labeled 'Learning Rate (log scale)' but the tick values are displayed linearly; please use a log-scaled axis consistently.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the benchmark compares learned classifiers on an external public dataset and makes no fitted-parameter-based prediction that feeds back into its own claim.

full rationale

The paper's central claim is an empirical benchmarking result: ViT-based and hybrid models are evaluated on a fixed, externally sourced CMS Open Data jet-image dataset (Andrews et al. [1]) and scored against CNN baselines using standard metrics. No equation in the paper is fitted to a subset of the target results and then reused to 'predict' those same results; the only preprocessing formulas (Eqs. 3-13) normalize inputs or define augmentation and are applied uniformly before model training, not used as evidence for the architecture ranking. Self-citation [19] (a Lorentz-equivariant quantum graph neural network paper by overlapping authors) appears only in the related-work list and is not load-bearing for any conclusion in this work. The abstract's 'consistently outperform' claim is broader than what Table 1 strictly supports, and the unequal application of Mixup to transformer models but not CNN baselines is a legitimate experimental-design concern about confounded comparison, not a circularity. There is no self-definitional step, no renamed known result, and no imported uniqueness theorem. The derivation chain is therefore self-contained with respect to circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper's claim depends on a set of domain assumptions about the simulation data, transfer learning, and augmentation physics, plus hand-chosen hyperparameters. No invented entities are introduced.

free parameters (5)
  • Mixup alpha = 0.2
    Beta(0.2,0.2) mixing coefficient used only for ViT/Swin models (Section 3.2, Eqs. 12-13); this asymmetric regularization confounds the ViT-vs-CNN comparison.
  • Initial learning rate = 1e-4 (head), 1e-6 (unfrozen blocks)
    Chosen in Section 3.4; sensitivity analysis shows 5e-5 is better, so the main benchmark may run at a suboptimal learning rate.
  • Weight decay = 1e-4
    Chosen in Section 3.4; sensitivity analysis finds 0.01 optimal, so the chosen value may not be ideal.
  • Dropout probability = 0.1 (ensemble head)
    Ensemble MLP head uses dropout delta=0.1 (Section 3.3); sensitivity analysis suggests 0.3 is more reliable.
  • Batch size = 32
    Used for all runs (Section 3.4); sensitivity analysis finds 64 better, so this choice may disadvantage all models compared to the optimum.
assumptions (4)
  • domain assumption The 2012 CMS Open Data simulated events and the derived 933,206-image dataset are a reliable proxy for quark-gluon jet classification under pileup conditions.
    The paper relies on Geant4 simulation and Pythia6 with Z2* tune (Section 3.1) rather than real collision data, and the conclusions are about detector-image classification generally.
  • domain assumption Pre-trained ImageNet weights transfer useful features to calorimeter jet images.
    All models use ImageNet-1k pretrained weights (Section 3.3), an implicit assumption that natural-image features help for detector images.
  • ad hoc to paper Random rotations, horizontal flips, resized cropping, and color jitter preserve the physics content of jet images.
    These augmentations (Section 3.2) are standard for natural images but are not physically justified for energy deposits in eta-phi space; color jitter in particular changes energy values arbitrarily.
  • ad hoc to paper The 7k-sample benchmark subset is representative of the full dataset.
    Table 1 is reported on 7k samples (Section 3.1 and Table 1) without specifying the sampling procedure, so the comparison assumes this subset is unbiased.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Vision Transformers for End-to-End Quark-Gluon Jet Classification from Calorimeter Images." pith.science (2026). https://pith.science/paper/K2Z27LYA

@misc{pith2026250614934,
  author       = {Pith},
  title        = {Pith review of: Vision Transformers for End-to-End Quark-Gluon Jet Classification from Calorimeter Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K2Z27LYA}},
  note         = {Machine review of arXiv:2506.14934}
}
read the original abstract

Distinguishing between quark- and gluon-initiated jets is a critical and challenging task in high-energy physics, pivotal for improving new physics searches and precision measurements at the Large Hadron Collider. While deep learning, particularly Convolutional Neural Networks (CNNs), has advanced jet tagging using image-based representations, the potential of Vision Transformer (ViT) architectures, renowned for modeling global contextual information, remains largely underexplored for direct calorimeter image analysis, especially under realistic detector and pileup conditions. This paper presents a systematic evaluation of ViTs and ViT-CNN hybrid models for quark-gluon jet classification using simulated 2012 CMS Open Data. We construct multi-channel jet-view images from detector-level energy deposits (ECAL, HCAL) and reconstructed tracks, enabling an end-to-end learning approach. Our comprehensive benchmarking demonstrates that ViT-based models, notably ViT+MaxViT and ViT+ConvNeXt hybrids, consistently outperform established CNN baselines in F1-score, ROC-AUC, and accuracy, highlighting the advantage of capturing long-range spatial correlations within jet substructure. This work establishes the first systematic framework and robust performance baselines for applying ViT architectures to calorimeter image-based jet classification using public collider data, alongside a structured dataset suitable for further deep learning research in this domain.

Figures

Figures reproduced from arXiv: 2506.14934 by the authors.

Figure 1
Figure 1. (a) Visualization of a representative gluon (left) and quark (right) jet event across the Tracks, ECAL, and HCAL detector channels. Logarithmic normalization is used for Tracks and ECAL, while linear scaling is applied to the sparse HCAL data. Colorbars indicate per-channel pixel intensities. (b) Average per-pixel intensity maps over N = 104 gluon (left) and quark (right) jets for each detector channel. Logarithmic … view at source ↗
Figure 2
Figure 2. Ablation study summarizing the effect of key training and architectural hyper￾parameters on model performance. Each subfigure isolates a single factor while holding others constant, illustrating its individual impact. ROC-AUC, with marked gains after unfreezing the ViT backbone at epoch 6 and the MaxViT backbone at epoch 10. The final epoch yielded the highest F1 score (0.6975) and ROC-AUC (0.7327), highlighting the… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 17 canonical work pages

  1. [22]

    Eye38(13), 2644–2648 (Sep 2024)

    KÄśran Yenice, E., Kara, C., Erdaŧ, Ã.B.: Automated detection of type 1 ROP, type 2 ROP and A-ROP based on deep learning. Eye38(13), 2644–2648 (Sep 2024). https://doi.org/10.1038/s41433-024-03184-0, https://www.nature.com/ articles/s41433-024-03184-0, publisher: Nature Publishing Group

    work page doi:10.1038/s41433-024-03184-0 2024

  2. [13]

    Advances in Neural Information Processing Systems34, 21056–21069 (2021), https://proceedings.neurips.cc/paper/2021/hash/ afe434653a898da20044041262b3ac74-Abstract.html

    Fang, W., Yu, Z., Chen, Y., Huang, T., Masquelier, T., Tian, Y.: Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems34, 21056–21069 (2021), https://proceedings.neurips.cc/paper/2021/hash/ afe434653a898da20044041262b3ac74-Abstract.html

    work page 2021

  3. [1]

    Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment977, 164304 (Oct 2020)

    Andrews, M., Alison, J., An, S., Burkle, B., Gleyzer, S., Narain, M., Paulini, M., Poczos, B., Usai, E.: End-to-end jet classification of quarks and gluons with the CMS Open Data. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment977, 164304 (Oct 2020). https://doi.org/10.1016/j.n...

    work page arXiv 2020

  4. [2]

    Computing and Software for Big Science4(1), 6 (Dec 2020)

    Andrews, M., Paulini, M., Gleyzer, S., Poczos, B.: End-to-End Physics Event Classifi- cationwithCMSOpenData:ApplyingImage-BasedDeepLearningtoDetectorData for the Direct Classification of Collision Events at the LHC. Computing and Software for Big Science4(1), 6 (Dec 2020). https://doi.org/10.1007/s41781-020-00038-8, http://link.springer.com/10.1007/s41781...

    work page doi:10.1007/s41781-020-00038-8 2020

  5. [3]

    Scientific Data9(1), 31 (Feb 2022)

    Chen, Y., Huerta, E.A., Duarte, J., Harris, P., Katz, D.S., Neubauer, M.S., Diaz, D., Mokhtar, F., Kansal, R., Park, S.E., Kindratenko, V.V., Zhao, Z., Rusack, R.: A FAIR and AI-ready Higgs boson decay dataset. Scientific Data9(1), 31 (Feb 2022). https://doi.org/10.1038/s41597-021-01109-0, https://www.nature.com/ articles/s41597-021-01109-0

    doi:10.1038/s41597-021-01109-0 2022

  6. [4]

    https://doi.org/10

    CMS Collaboration: GJet_Pt40_doubleEMEnriched_TuneZ2star_8TeV_ext- pythia6 in AODSIM format for 2012 collision data (2017). https://doi.org/10. 7483/OPENDATA.CMS.2W51.W8AT, http://opendata.cern.ch/record/7778, uRL: http://opendata.cern.ch/record/7778

    work page 2017

  7. [5]

    https://doi.org/10.1088/1748-0221/3/ 08/S08004, https://iopscience.iop.org/article/10.1088/1748-0221/3/08/S08004

    CMSCollaborationetal.:TheCMSexperimentattheCERNLHC.JournalofInstru- mentation3(08), S08004–S08004 (Aug 2008). https://doi.org/10.1088/1748-0221/3/ 08/S08004, https://iopscience.iop.org/article/10.1088/1748-0221/3/08/S08004

    doi:10.1088/1748-0221/3/ 2008

  8. [6]

    Physical Review Letters131(6), 061801 (Aug 2023)

    CMSCollaborationetal.:SearchforHiggsBosonDecaytoaCharmQuark-Antiquark Pair in Proton-Proton Collisions at $\sqrt{s}=13\text{ }\text{ }\mathrm{TeV}$. Physical Review Letters131(6), 061801 (Aug 2023). https://doi.org/10. 1103/PhysRevLett.131.061801, https://link.aps.org/doi/10.1103/PhysRevLett.131. 061801, publisher: American Physical Society

    doi:10.1103/physrevlett.131 2023

Show all 36 references
  1. [7]

    Physical Review Letters131(4), 041803 (Jul 2023)

    CMSCollaborationetal.:SearchforNonresonantPairProductionofHighlyEnergetic Higgs Bosons Decaying to Bottom Quarks. Physical Review Letters131(4), 041803 (Jul 2023). https://doi.org/10.1103/PhysRevLett.131.041803, https://link.aps.org/ doi/10.1103/PhysRevLett.131.041803

    2023 doi

  2. [8]

    Advances in neural information processing sys- tems34, 3965–3977 (2021), https://proceedings.neurips.cc/paper/2021/hash/ 20568692db622456cc42a2e853ca21f8-Abstract.html

    Dai, Z., Liu, H., Le, Q.V., Tan, M.: Coatnet: Marrying convolution and at- tention for all data sizes. Advances in neural information processing sys- tems34, 3965–3977 (2021), https://proceedings.neurips.cc/paper/2021/hash/ 20568692db622456cc42a2e853ca21f8-Abstract.html

    2021

  3. [9]

    Journal of High Energy Physics2016(7), 69 (Jul 2016)

    De Oliveira, L., Kagan, M., Mackey, L., Nachman, B., Schwartzman, A.: Jet- images — deep learning edition. Journal of High Energy Physics2016(7), 69 (Jul 2016). https://doi.org/10.1007/JHEP07(2016)069, http://link.springer.com/10. 1007/JHEP07(2016)069

    2016 doi

  4. [10]

    SciPost Physics12(6), 188 (Jun 2022)

    Dillon, B.M., Kasieczka, G., Olischlager, H., Plehn, T., Sorrenson, P., Vogel, L.: Symmetries, Safety, and Self-Supervision. SciPost Physics12(6), 188 (Jun 2022). https://doi.org/10.21468/SciPostPhys.12.6.188, http://arxiv.org/abs/2108.04253, arXiv:2108.04253 [hep-ph]

    2022 arXiv

  5. [11]

    arXiv preprint arXiv:2010.11929 (2020), https://arxiv.org/pdf/2010.11929/1000

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020), https://arxiv.org/pdf...

    2020 arXiv

  6. [12]

    The European Physical Journal C82(6), 564 (Jun 2022)

    Dreyer, F.A., Grabarczyk, R., Monni, P.F.: Leveraging universality of jet taggers through transfer learning. The European Physical Journal C82(6), 564 (Jun 2022). https://doi.org/10.1140/epjc/s10052-022-10469-9, https://doi.org/10.1140/epjc/ s10052-022-10469-9

    2022 doi

  7. [14]

    Journal of High Energy Physics2024(7), 146 (Jul 2024)

    Furuichi, A., Lim, S.H., Nojiri, M.M.: Jet classification using high-level features from anatomy of top jets. Journal of High Energy Physics2024(7), 146 (Jul 2024). https: //doi.org/10.1007/JHEP07(2024)146, https://doi.org/10.1007/JHEP07(2024)146

    2024 doi

  8. [15]

    Journal of High Energy Physics2022(7), 30 (Jul 2022)

    Gong, S., Meng, Q., Zhang, J., Qu, H., Li, C., Qian, S., Du, W., Ma, Z.M., Liu, T.Y.: An efficient Lorentz equivariant graph neural network for jet tagging. Journal of High Energy Physics2022(7), 30 (Jul 2022). https://doi.org/10.1007/JHEP07(2022)030, https://doi.org/10.1007/J...

    2022 doi

  9. [16]

    Physical Review D 94(11), 112002 (Dec 2016)

    Guest, D., Collado, J., Baldi, P., Hsu, S.C., Urban, G., Whiteson, D.: Jet flavor classification in high-energy physics with deep neural networks. Physical Review D 94(11), 112002 (Dec 2016). https://doi.org/10.1103/PhysRevD.94.112002, https: //link.aps.org/doi/10.1103/PhysRev...

    2016 doi

  10. [17]

    Computational Visual Media7(2), 187–199 (Jun 2021)

    Guo, M.H., Cai, J.X., Liu, Z.N., Mu, T.J., Martin, R.R., Hu, S.M.: PCT: Point cloud transformer. Computational Visual Media7(2), 187–199 (Jun 2021). https://doi. org/10.1007/s41095-021-0229-5, https://ieeexplore.ieee.org/document/10897555/

    2021

  11. [18]

    In: Deep Learning for Physical Sci- ences Workshop at the 31st Conference on Neural Information Processing Systems (NeurIPS) (2017)

    Henrion, I., Brehmer, J., Bruna, J., Cho, K., Cranmer, K., Louppe, G., Rochette, G.: Neural Message Passing for Jet Physics. In: Deep Learning for Physical Sci- ences Workshop at the 31st Conference on Neural Information Processing Systems (NeurIPS) (2017)

    2017

  12. [19]

    IEEE Transactions on Artificial Intelligence pp

    Jahin, M.A., Masud, M.A., Suva, M.W., Mridha, M.F., Dey, N.: Lorentz-Equivariant Quantum Graph Neural Network for High-Energy Physics. IEEE Transactions on Artificial Intelligence pp. 1–11 (2025). https://doi.org/10.1109/TAI.2025.3554461, https://ieeexplore.ieee.org/document/10938398

    2025

  13. [20]

    https://doi.org/10.1038/s41586-021-03819-2, https://www.nature.com/articles/s41586-021-03819-2

    Jumper,Johnetal.:HighlyaccurateproteinstructurepredictionwithAlphaFold.Na- ture596(7873), 583–589 (Aug 2021). https://doi.org/10.1038/s41586-021-03819-2, https://www.nature.com/articles/s41586-021-03819-2

    2021 doi

  14. [21]

    Journal of High Energy Physics2019(1), 121 (Jan 2019)

    Komiske,P.T.,Metodiev,E.M.,Thaler,J.:Energyflownetworks:deepsetsforparticle jets. Journal of High Energy Physics2019(1), 121 (Jan 2019). https://doi.org/10. 1007/JHEP01(2019)121, https://link.springer.com/10.1007/JHEP01(2019)121

    2019 doi

  15. [23]

    Physics Reports841, 1– 63 (2020), https://www.sciencedirect.com/science/article/pii/S0370157319303643, publisher: Elsevier

    Larkoski,A.J.,Moult,I.,Nachman,B.:JetsubstructureattheLargeHadronCollider: a review of recent advances in theory and machine learning. Physics Reports841, 1– 63 (2020), https://www.sciencedirect.com/science/article/pii/S0370157319303643, publisher: Elsevier

    2020

  16. [24]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012–10022 (2021), https://openaccess.thecvf.com/...

    2021

  17. [25]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976–11986 (2022), http://openaccess.thecvf.com/content/ CVPR2022/html/Liu_A_ConvNet_for_th...

    2022

  18. [26]

    Journal of High Energy Physics2019(1), 57 (Jan 2019)

    Louppe,G.,Cho,K.,Becot,C.,Cranmer,K.:QCD-awarerecursiveneuralnetworksfor jet physics. Journal of High Energy Physics2019(1), 57 (Jan 2019). https://doi.org/ 10.1007/JHEP01(2019)057, https://link.springer.com/10.1007/JHEP01(2019)057

    2019 doi

  19. [27]

    Machine Learning: Science and Technology2(3), 035027 (Sep 2021)

    Mikuni, V., Canelli, F.: Point cloud transformers applied to collider physics. Machine Learning: Science and Technology2(3), 035027 (Sep 2021). https://doi.org/10.1088/ 2632-2153/ac07f6, https://iopscience.iop.org/article/10.1088/2632-2153/ac07f6

    2021 doi

  20. [28]

    Mikuni, V., Nachman, B.: Method to simultaneously facilitate all jet physics tasks. Phys. Rev. D111, 054015 (Mar 2025). https://doi.org/10.1103/PhysRevD.111. 054015, https://link.aps.org/doi/10.1103/PhysRevD.111.054015

    2025 doi

  21. [29]

    Physical Review D101(5), 056019 (Mar 2020)

    Qu, H., Gouskos, L.: Jet tagging via particle clouds. Physical Review D101(5), 056019 (Mar 2020). https://doi.org/10.1103/PhysRevD.101.056019, https://link. aps.org/doi/10.1103/PhysRevD.101.056019

    2020 doi

  22. [30]

    et al.: Particle-flow reconstruction and global event description with the cms detector

    Sirunyan, A.M. et al.: Particle-flow reconstruction and global event description with the cms detector. Journal of Instrumentation12(10), P10003 (oct 2017). https://doi. org/10.1088/1748-0221/12/10/P10003, https://dx.doi.org/10.1088/1748-0221/12/ 10/P10003

    2017 doi

  23. [31]

    Journal of Instrumentation15(06), P06005– P06005 (Jun 2020)

    Sirunyan,A.M.etal.:Identificationofheavy,energetic,hadronicallydecayingparticles using machine-learning techniques. Journal of Instrumentation15(06), P06005– P06005 (Jun 2020). https://doi.org/10.1088/1748-0221/15/06/P06005, https:// iopscience.iop.org/article/10.1088/1748-022...

    2020 doi

  24. [32]

    In: International conference on machine learning

    Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105–6114. PMLR (2019), http://proceedings.mlr.press/v97/tan19a.html?ref=jina-ai-gmbh.ghost.io

    2019

  25. [33]

    In: Avidan, S., Brostow, G., CissÃľ, M., Farinella, G.M., Hassner,T.(eds.)ComputerVisionâĂŞECCV2022,vol.13684,pp.459–479.Springer Nature Switzerland, Cham (2022)

    Tu,Z.,Talebi,H.,Zhang,H.,Yang,F.,Milanfar,P.,Bovik,A.,Li,Y.:MaxViT:Multi- axis Vision Transformer. In: Avidan, S., Brostow, G., CissÃľ, M., Farinella, G.M., Hassner,T.(eds.)ComputerVisionâĂŞECCV2022,vol.13684,pp.459–479.Springer Nature Switzerland, Cham (2022). https://doi.org...

    2022 doi

  26. [34]

    Advances in neural information processing systems30(2017), https://proceedings.neurips.cc/paper/2017/hash/ 3f5ee243547dee91fbd053c1c4a845aa-Abstract.html

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Å., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems30(2017), https://proceedings.neurips.cc/paper/2017/hash/ 3f5ee243547dee91fbd053c1c4a845aa-Abstract.html

    2017

  27. [35]

    ACM Transactions on Graphics38(5), 1–12 (Oct 2019)

    Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic Graph CNN for Learning on Point Clouds. ACM Transactions on Graphics38(5), 1–12 (Oct 2019). https://doi.org/10.1145/3326362, https://dl.acm.org/doi/10.1145/ 3326362

    2019 doi

  28. [36]

    https://doi.org/10.48550/arXiv.2110.00476, http: //arxiv.org/abs/2110.00476, arXiv:2110.00476 [cs]

    Wightman, R., Touvron, H., JÃľgou, H.: ResNet strikes back: An improved training procedure in timm (Oct 2021). https://doi.org/10.48550/arXiv.2110.00476, http: //arxiv.org/abs/2110.00476, arXiv:2110.00476 [cs]

    2021 doi arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.