Pith. sign in

REVIEW 3 major objections 7 minor 63 references

A Transformer on raw calorimeter cells separates collimated diphoton jets from single photons far better than shower-shape methods and also regresses the diphoton mass.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-10 11:51 UTC pith:FAJQVE2Q

load-bearing objection Solid simulation methods paper: cell-level Transformer clearly beats SSV and other cell models on a hard collimation regime, with usable mass regression and public code; main limit is idealized GEANT4 transfer. the 3 major comments →

arxiv 2607.08175 v1 pith:FAJQVE2Q submitted 2026-07-09 hep-ph hep-ex

Transformer-based machine learning using low-level calorimeter signals for collimated photon identification at collider experiments

Gabriel Matos , Lauren Larson , Abhilasha Dave , Maria Bressan , Azal Amer , Cindy Liu , Nikiforos Nikiforou , Jonathan Long
show 3 more authors
Timothy Andeen John Parsons Julia Gonski
This is my paper
classification hep-ph hep-ex
keywords calorimeter cell-level learningTransformerphoton-jet classificationaxion-like particlesinvariant-mass regressionMLP Mixershower shape variablescollider trigger
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Collider calorimeters record energy deposit cell by cell, but standard photon identification throws most of that detail away and works with a handful of summary shower-shape numbers. This paper shows that feeding the cells themselves—up to the 200 highest-energy ones, each tagged by energy, position and layer—into modern machine-learning models yields a much sharper separation between true single photons and the highly collimated photon pairs that come from light axion-like particles. Among six architectures tested on a GEANT4 ATLAS-like simulation, a multi-task Transformer reaches an overall AUC of 0.98 and background rejection roughly two orders of magnitude better than the shower-shape baselines at 90 % signal efficiency. The same network also reconstructs the diphoton invariant mass directly from the cells with a resolution of about 64 MeV, cleanly resolving the π⁰ and η peaks that are the main fake-photon backgrounds. An MLP Mixer, while slightly less powerful, uses only 3 % of the compute and is therefore a candidate for real-time trigger use. The work therefore argues that low-level cell information, once properly modelled, can open regions of BSM parameter space that conventional high-level reconstruction cannot reach.

Core claim

Cell-level machine learning, and in particular a multi-task Transformer that attends over the energy-ordered top-200 calorimeter cells, classifies highly collimated ALP→γγ photon-jets from single photons with an overall test AUC of 0.98 and background rejection of order 470 at 90 % signal efficiency—far above shower-shape BDT/DNN and other cell-level baselines—while the same network regresses the diphoton mass to ~64 MeV resolution, resolving π⁰ and η fakes.

What carries the argument

A Transformer encoder with multi-head self-attention over variable-length cell tokens (energy, transverse coordinates relative to the shower barycentre, and layer), followed by global attention pooling and two task heads that jointly optimise binary classification and LogCosh mass regression.

Load-bearing premise

That the cell-level correlations the Transformer exploits will still be present and useful once the model is moved from an idealised flat-face GEANT4 calorimeter without pile-up, magnetic field or electronics noise onto real collider data and trigger hardware.

What would settle it

Train and evaluate the identical cell-level Transformer on full ATLAS or CMS simulation (or real data) that includes pile-up, magnetic field, dead-material non-uniformities and realistic clustering; if the AUC advantage over shower-shape baselines collapses below ~0.95 or the mass resolution degrades beyond ~100 MeV, the claimed gains do not transfer.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The manuscript studies machine-learning classification of highly collimated ALP→γγ photon-jets versus isolated single photons using an ATLAS-like GEANT4 calorimeter model. It compares six architectures: SSV-based BDT and DNN baselines, and cell-level CNN, PFN, Transformer, and MLP Mixer models operating on up to 200 energy-ordered cells (E, x, y, layer). The multi-task Transformer achieves the best inclusive test AUC (0.98) and the strongest background rejection at fixed signal efficiency (Table 3: ~472 at 90% ε_sig), retains useful discrimination down to O(10) MeV ALP masses, and simultaneously regresses the diphoton invariant mass with ~64 MeV resolution, resolving π0 and η peaks. The MLP Mixer is presented as a lower-cost alternative (~3% of Transformer FLOPs) for potential trigger use. Code and samples are released.

Significance. If the reported gains transfer even partially to real reconstruction and trigger environments, the work would meaningfully improve photon identification and sensitivity to light ALP and related photon-jet signatures in a regime (ΔR_γγ ≲ calorimeter cell size) that has been difficult for conventional SSV methods. The systematic six-architecture comparison, mass-binned ROCs, AUC vs (m_a, p_T) maps, multi-seed uncertainties on the top models, and public code/data are clear strengths. The multi-task mass regression from cells is a useful additional handle for π0/η rejection. The main caveat is that all results are obtained in an idealized particle-gun simulation without pileup, electronics noise, magnetic field, or full topological clustering; the authors acknowledge this, but the broader impact claims rest on transfer that is not yet demonstrated.

major comments (3)
  1. Abstract and §5 claim that cell-level methods “can extend calorimeter-based particle identification … well beyond the capabilities of current conventional techniques.” The supporting evidence (Table 3, Figs. 6–13) is obtained in a highly idealized GEANT4 setup (§2: flat face at η=0, uniform 2.5 X0 Al dead material, no B-field, no pileup, no electronics noise, particle-gun incidence, energy-ordered top-200 cells without full topo-clustering). This limitation is stated but under-weighted relative to the impact language. Please either (i) qualify the abstract/conclusions more tightly to the simulation scope, or (ii) add a short quantitative discussion of which effects (pileup, noise, clustering thresholds, non-uniform material) are most likely to degrade the pairwise cell correlations the Transformer exploits, and what validation path is planned.
  2. §3.1 / Table 3 comparison fairness: the CNN is restricted to the first three EM layers and a multi-task objective, while the PFN, Transformer, and MLP Mixer use all six layers (including HCAL). The text notes that remaining layers “contribute no improvement” for the CNN, but this is not shown and the input dimensionality therefore differs across cell-level models. For a load-bearing ranking claim, either retrain the CNN with the same six-layer cell set (or show an ablation that HCAL adds nothing for all architectures) or clearly mark the CNN as a reduced-input baseline so the Table 3 ordering is not over-interpreted.
  3. §4.1 mass regression and π0/η handle: the Transformer reconstructs m_π0 ≈ 172 MeV and m_η ≈ 564 MeV (Fig. 12) with an upward bias comparable to the ~60 MeV resolution, and a small non-zero m_γ bias. The abstract and conclusions present this as “an additional handle in reducing the π0 and η fake photon backgrounds,” but no background-rejection or purity numbers versus a pure classifier cut (or versus SSV-based π0 rejection) are given. A short quantification—e.g., rejection of π0/η samples at fixed single-photon efficiency using the mass head, or a combined score+mass working point—would make this claim load-bearing rather than qualitative.
minor comments (7)
  1. Eq. (2.1) and surrounding text: ΔR_γγ ≈ 2 m_a / p_T,a is the ultra-collinear approximation; a one-line statement of its validity range for the generated sample would help readers.
  2. §3.2: the multi-task loss weight L = L_classifier + 4 L_regression is “chosen empirically.” A brief sensitivity check (or statement that AUC is stable under O(1) changes of the weight) would strengthen reproducibility.
  3. Fig. 8 / Fig. 15: the ΔR_γγ bands are useful; adding the finest cell angular scale (~3×10^{-3}) as a horizontal reference line on the AUC map would make the “sub-cell” regime more immediate.
  4. Table 2: FLOPs/MACs are quoted for a single inference pass; stating the assumed batch size and whether padding/masking is included would aid FPGA-oriented readers.
  5. Appendix C / Fig. 17: SSV definitions are said to match ATLAS photon ID plus E_max_cell; a compact table of exact formulas (or a pointer to the ATLAS note equations) would remove any ambiguity for reimplementation.
  6. Typographical: “V ariables” in the Contents (C Shower Shape V ariables) and occasional spacing issues around π0/η in the abstract PDF rendering.
  7. §4.3: background rejection for BDT/DNN/CNN/MLP Mixer is given without multi-seed uncertainties, unlike Transformer/PFN. Even a single-seed note or a short stability check would make Table 3 more uniform.

Circularity Check

0 steps flagged

No significant circularity: standard supervised multi-task ML on independent GEANT4 truth labels with held-out test metrics.

full rationale

The paper's central claims are empirical performance numbers (Transformer test AUC 0.98, background rejection ~472 at 90% signal efficiency, mass resolution 64 MeV) obtained by training classifiers/regressors on simulated calorimeter cells whose labels (ALP vs single-photon; true m_a) are generated independently by GEANT4 particle-gun events. Train/validation/test splits are random 80:10:10; ROC/AUC and residual distributions are evaluated on held-out events and are not algebraically forced by the loss (cross-entropy + LogCosh) or by any fitted constant. Architecture comparisons (BDT/DNN on SSVs vs CNN/PFN/Transformer/MLP-Mixer on cells) are ablations under identical simulation conditions; none of the reported gains reduce by construction to an input definition or a self-cited uniqueness theorem. The sole self-reference is the authors' Zenodo GEANT4 geometry package used to produce the public datasets; this is infrastructure for reproducibility, not a load-bearing premise that the classification or mass-regression results are true by definition. No self-definitional loops, fitted-input-as-prediction, or ansatz-smuggling via citation appear in the derivation chain. The idealized-simulation transfer caveat is an external validity issue, not circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim rests on standard ML practice plus domain modeling of an ATLAS-like calorimeter and ALP kinematics. No new physical entity is postulated. Free parameters are ordinary architecture/training choices and simulation ranges; domain assumptions about GEANT4 fidelity and the simplified detector are the load-bearing external premises.

free parameters (6)
  • Classification vs regression loss weight (Transformer) = 4
    L = L_classifier + 4 L_regression chosen empirically so both losses contribute roughly equally (§3.2); affects multi-task performance.
  • CNN multi-task loss weights = 0.8 / 0.2
    L = 0.8 L_classifier + 0.2 L_regression chosen for the CNN multi-task objective (§3.1).
  • Max cells per event N_cells = 200
    Cap of 200 energy-ordered cells per event, motivated by ATLAS topo-cluster occupancy; truncates/pads inputs for all cell-level models (§3).
  • Transformer architecture hyperparameters = as listed in §3.2
    Embedding 128, 4 encoder layers, 8 heads, FFN 256, dropout 0.065, task heads 128-64-32 selected via hyperparameter scan (§3.2).
  • MLP Mixer embedding and mixer width = 50
    Patch embedding dimension 50 and single mixer block width 50 chosen for lightweight design (§3.3).
  • Simulated ALP mass and p_T ranges = 10 MeV–2.5 GeV; 50–300 GeV
    Uniform coverage m_a ∈ [10 MeV, 2.5 GeV], p_T ∈ [50, 300] GeV defines the collimation regime and evaluation bins (§2).
axioms (5)
  • domain assumption GEANT4 shower simulation with the stated ATLAS-like layered geometry and 2.5 X0 aluminum dead material adequately represents EM shower differences between single photons and collimated diphotons for ranking ML methods.
    Entire training and evaluation dataset is generated this way (§2, Fig. 2); no real collision data.
  • domain assumption Energy-ordered selection of at most 200 cells without full topological clustering preserves the discriminating cell-level information.
    Input construction for all cell-level models (§3).
  • domain assumption ALP→γγ kinematics with ΔR_γγ ≈ 2 m_a / p_T,a and prompt decay produce the target photon-jet topology.
    Signal definition and collimation bands in §2, Eq. (2.1).
  • standard math Standard ML training assumptions: i.i.d. train/val/test split, binary cross-entropy / LogCosh objectives, Adam/AdamW optimization converge to useful minima without pathological label leakage.
    Training protocols in §3.1–3.3.
  • domain assumption Shower-shape variables defined as in ATLAS photon ID (plus E_max_cell) are a fair high-level baseline for comparison.
    Appendix C and BDT/DNN inputs in §3.1.

pith-pipeline@v1.1.0-grok45 · 23946 in / 3682 out tokens · 42810 ms · 2026-07-10T11:51:31.194374+00:00 · methodology

0 comments
read the original abstract

Electromagnetic calorimeters provide essential information for reconstructing and selecting both Standard Model (SM) and potential beyond the SM physics events at high-energy particle colliders. The fine-grained segmentation of modern calorimeters captures rich information about the internal structure of particle showers, much of which is discarded by conventional high-level reconstruction methods. In this work, we leverage calorimeter cell-level information to classify highly collimated diphoton signatures, arising from the decay of light axion-like particles, from isolated single-photon showers. We systematically compare a range of machine learning architectures, spanning high-level, shower shape variable-based approaches and direct cell-level methods. Cell-level machine learning shows significantly superior classification ability, with a Transformer in particular representing the best performance among six different architectures studied, and an MLP Mixer representing a resource-constrained alternative for potential real-time, trigger-level applications. Beyond classification, the Transformer model developed enables direct invariant mass regression from calorimeter cells, improving the characterization of light resonances and providing an additional handle in reducing the $\pi^0$ and $\eta$ fake photon backgrounds. These results demonstrate that cell-level machine learning methods can extend calorimeter-based particle identification and performance well beyond the capabilities of current conventional techniques.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

63 extracted references · 63 canonical work pages · 30 internal anchors

  1. [1]

    ATLAS collaboration,ATLAS Liquid Argon Calorimeter: Technical Design Report, Tech. Rep. CERN-LHCC-96-041, CERN, Geneva (1996), DOI

    work page 1996

  2. [2]

    ATLAS collaboration,The ATLAS Experiment at the CERN Large Hadron Collider,JINST 3(2008) S08003

    work page 2008

  3. [3]

    Evans and P

    L. Evans and P. Bryant,LHC Machine,JINST3(2008) S08001

    work page 2008

  4. [4]

    Photons, Photon Jets and Dark Photons at 750 GeV and Beyond

    B. Dasgupta, J. Kopp and P. Schwaller,Photons, photon jets, and dark photons at 750 GeV and beyond,Eur. Phys. J. C76(2016) [1602.04692]

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  5. [5]

    Diphotons from Diaxions

    L. Aparicio, A. Azatov, E. Hardy and A. Romanino,Diphotons from diaxions,JHEP2016 (2016) [1602.00949]

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  6. [6]

    Rays of light from the LHC

    S. Knapen, T. Melia, M. Papucci and K.M. Zurek,Rays of light from the LHC,Phys. Rev. D 93(2016) [1512.04928]

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  7. [7]

    Interpreting the 750 GeV Di-photon Resonance using photon-jets in Hidden-Valley-like models

    J. Chang, K. Cheung and C.-T. Lu,Interpreting the 750 GeV diphoton resonance using photon jets in hidden-valley-like models,Phys. Rev. D93(2016) [1512.06671]

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  8. [8]

    Peccei and H.R

    R.D. Peccei and H.R. Quinn,Constraints Imposed by CP Conservation in the Presence of Instantons,Phys. Rev. D16(1977) 1791

    work page 1977

  9. [9]

    Peccei and H.R

    R.D. Peccei and H.R. Quinn,CP Conservation in the Presence of Pseudoparticles,Phys. Rev. Lett.38(1977) 1440

    work page 1977

  10. [10]

    Wilczek,Problem of StrongPandTInvariance in the Presence of Instantons,Phys

    F. Wilczek,Problem of StrongPandTInvariance in the Presence of Instantons,Phys. Rev. Lett.40(1978) 279

    work page 1978

  11. [11]

    Weinberg,A New Light Boson?,Phys

    S. Weinberg,A New Light Boson?,Phys. Rev. Lett.40(1978) 223

    work page 1978

  12. [12]

    ALPs at Colliders

    K. Mimasu and V. Sanz,ALPs at Colliders,JHEP06(2015) 173 [1409.4792]

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  13. [13]

    Collider Probes of Axion-Like Particles

    M. Bauer, M. Neubert and A. Thamm,Collider probes of axion-like particles,JHEP2017 (2017) [1708.00443]

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  14. [14]

    Feebly-Interacting Particles:FIPs 2020 Workshop Report

    P. Agrawal, M. Bauer, J. Beacham, A. Berlin, A. Boyarsky, S. Cebrian et al., Feebly-interacting particles: FIPs 2020 workshop report,Eur. Phys. J. C81(2021) [2102.12143]

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  15. [15]

    Axion-like Particle Search with a Light-Shining-Through-Walls Setup at a $\gamma$-$\gamma$ Collider

    Z.-Y. Yan and J. Feng,Axion-like Particle Search with a Light-Shining-Through-Walls Setup at aγ-γCollider,2512.15192

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    Lentz, C.R

    E.W. Lentz, C.R. Boutan, M.S. Taubman and K.L. Gervais,Developing centimeter-scale-cavity arrays for axion dark matter detection in the 100 micro-electron-volt range,JINST21(2026) P04020 [2601.21074]

    work page arXiv 2026

  17. [17]

    Akg¨ um¨ us, N

    M.A. Akg¨ um¨ us, N. Salama, J. Egge, E. Garutti, M. Maroudas, L.H. Nguyen et al.,A new limit for axion dark matter with SPACE,JCAP04(2026) 054 [2506.18411]

    work page arXiv 2026

  18. [18]

    Axion Dark Matter: What is it and Why Now?

    F. Chadha-Day, J. Ellis and D.J.E. Marsh,Axion dark matter: What is it and why now?, Sci. Adv.8(2022) [2105.01406]

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  19. [19]

    Axion Cosmology

    D.J.E. Marsh,Axion Cosmology,Phys. Rept.643(2016) 1 [1510.07633]

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  20. [20]

    Dine and W

    M. Dine and W. Fischler,The Not So Harmless Axion,Phys. Lett. B120(1983) 137

    work page 1983

  21. [21]

    Abbott and P

    L.F. Abbott and P. Sikivie,A Cosmological Bound on the Invisible Axion,Phys. Lett. B120 (1983) 133. – 23 –

    work page 1983

  22. [22]

    Preskill, M.B

    J. Preskill, M.B. Wise and F. Wilczek,Cosmology of the Invisible Axion,Phys. Lett. B120 (1983) 127

    work page 1983

  23. [23]

    Jaeckel and M

    J. Jaeckel and M. Spannowsky,Probing MeV to 90 GeV axion-like particles with LEP and LHC,Phys. Lett. B753(2016) 482–487

    work page 2016

  24. [24]

    A 750 GeV Diphoton Signal from a Very Light Pseudoscalar in the NMSSM

    U. Ellwanger and C. Hugonie,A 750 GeV Diphoton Signal from a Very Light Pseudoscalar in the NMSSM,JHEP(2016) 114 [1602.03344]

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  25. [25]

    CMS collaboration,Performance of photon reconstruction and identification with the CMS detector in proton-proton collisions at √s= 8 TeV,JINST10(2015) P08010–P08010

    work page 2015

  26. [26]

    ATLAS collaboration,Photon identification in 2015 ATLAS data, Tech. Rep. ATL-PHYS-PUB-2016-014, CERN, Geneva (2016)

    work page 2015

  27. [27]

    CMS collaboration,Electron and photon reconstruction and identification with the CMS experiment at the CERN LHC,JINST16(2021) P05014

    work page 2021

  28. [28]

    ATLAS collaboration,Measurement of the photon identification efficiencies with the ATLAS detector using LHC Run 2 data collected in 2015 and 2016,Eur. Phys. J. C79(2019) [1810.05087]

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  29. [29]

    ATLAS collaboration,Identification of electrons using a deep neural network in the ATLAS experiment, Tech. Rep. ATL-PHYS-PUB-2022-022, CERN, Geneva (2022)

    work page 2022

  30. [30]

    ATLAS collaboration,Search for pairs of highly collimated photon-jets inppcollisions at√s= 13TeV with the ATLAS detector,Phys. Rev. D99(2019) 012008

    work page 2019

  31. [31]

    ATLAS collaboration,Search for short- and long-lived axion-like particles inH→aa→4γ decays with the ATLAS experiment at the LHC,Eur. Phys. J. C84(2024) [2312.03306]

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  32. [32]

    ATLAS collaboration,Electron Identification with a Convolutional Neural Network in the ATLAS Experiment, Tech. Rep. ATL-PHYS-PUB-2023-001, CERN, Geneva (2023)

    work page 2023

  33. [33]

    CMS collaboration,Search for New Resonances Decaying to Pairs of Merged Diphotons in Proton-Proton Collisions at √s= 13 TeV,Phys. Rev. Lett.134(2025) 041801

    work page 2025

  34. [34]

    CMS collaboration,Search for Exotic Higgs Boson DecaysH→ AA →4γwith Events Containing Two Merged Diphotons in Proton-Proton Collisions at √s= 13 TeV,Phys. Rev. Lett.131(2023) 101801

    work page 2023

  35. [35]

    CMS collaboration,Reconstruction of decays to merged photons using end-to-end deep learning with domain continuation in the CMS detector,Phys. Rev. D108(2023) [2204.12313]

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  36. [36]

    CMS collaboration,Search for exotic Higgs boson decays H→ AAwithAA →γγin events with a semi-merged topology in proton-proton collisions at √s= 13 TeV,2601.00183

    work page arXiv

  37. [37]

    Energy Flow Networks: Deep Sets for Particle Jets

    P.T. Komiske, E.M. Metodiev and J. Thaler,Energy flow networks: deep sets for particle jets,JHEP2019(2019) [1810.05165]

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  38. [38]

    X. Ai, W.Y. Feng, S.-C. Hsu, K. Li and C.-T. Lu,Detecting highly collimated photon-jets from Higgs boson exotic decays with deep learning,2401.15690

    work page internal anchor Pith review Pith/arXiv arXiv

  39. [39]

    Attention Is All You Need

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A.N. Gomez et al.,Attention Is All You Need,1706.03762

    work page internal anchor Pith review Pith/arXiv arXiv

  40. [40]

    ATLAS collaboration,Graph Neural Network Jet Flavour Tagging with the ATLAS Detector, Tech. Rep. ATL-PHYS-PUB-2022-027, CERN, Geneva (2022). – 24 –

    work page 2022

  41. [41]

    ATLAS collaboration,Transforming jet flavour tagging at ATLAS,2505.19689

    work page arXiv

  42. [42]

    MLP-Mixer: An all-MLP Architecture for Vision

    I. Tolstikhin, N. Houlsby, A. Kolesnikov, L. Beyer, X. Zhai, T. Unterthiner et al., MLP-Mixer: An all-MLP Architecture for Vision,2105.01601

    work page internal anchor Pith review Pith/arXiv arXiv

  43. [43]

    ATLAS collaboration,Technical Design Report for the Phase-II Upgrade of the ATLAS TDAQ System, Tech. Rep. CERN-LHCC-2017-020, ATLAS-TDR-029, CERN, Geneva (2017), DOI

    work page 2017

  44. [44]

    Agostinelli et al.,Geant4– a simulation toolkit,Nucl

    S. Agostinelli et al.,Geant4– a simulation toolkit,Nucl. Instrum. Meth. A506(2003) 250

    work page 2003

  45. [45]

    gabrielpmatos/g4-atlas-calorimeter: Datasets for ML Trainings

    M. Pitt and G. Matos, “gabrielpmatos/g4-atlas-calorimeter: Datasets for ML Trainings.” Zenodo, 2026. https://zenodo.org/doi/10.5281/zenodo.18825682

    work page doi:10.5281/zenodo.18825682 2026

  46. [46]

    ATLAS collaboration,Electron and photon energy calibration with the ATLAS detector using LHC Run 2 data,JINST19(2024) P02009 [2309.05471]

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  47. [47]

    ATLAS collaboration,Electron reconstruction and identification in the ATLAS experiment using the 2015 and 2016 LHC proton–proton collision data at √s= 13TeV,Eur. Phys. J. C 79(2019) 639 [1902.04655]

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  48. [48]

    ATLAS collaboration,Topological cell clustering in the ATLAS calorimeters and its performance in LHC Run 1,Eur. Phys. J. C77(2017) 490 [1603.02934]

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  49. [49]

    Chen and C

    T. Chen and C. Guestrin,XGBoost: A Scalable Tree Boosting System, inProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16, p. 785–794, ACM, Aug., 2016, DOI

    work page 2016

  50. [50]

    PyTorch: An Imperative Style, High-Performance Deep Learning Library

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan et al.,PyTorch: An Imperative Style, High-Performance Deep Learning Library,1912.01703

    work page internal anchor Pith review Pith/arXiv arXiv 1912

  51. [51]

    Adam: A Method for Stochastic Optimization

    D.P. Kingma and J. Ba,Adam: A Method for Stochastic Optimization,1412.6980

    work page internal anchor Pith review Pith/arXiv arXiv

  52. [52]

    Spconv: Spatially Sparse Convolution Library

    Spconv Contributors, “Spconv: Spatially Sparse Convolution Library.” https://github.com/traveller59/spconv, 2022

    work page 2022

  53. [53]

    Submanifold Sparse Convolutional Networks

    B. Graham and L. van der Maaten,Submanifold Sparse Convolutional Networks, 1706.01307

    work page internal anchor Pith review Pith/arXiv arXiv

  54. [54]

    Chollet et al., “Keras.”https://keras.io, 2015

    F. Chollet et al., “Keras.”https://keras.io, 2015

    work page 2015

  55. [55]

    TensorFlow: A system for large-scale machine learning

    M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean et al.,TensorFlow: A system for large-scale machine learning,1605.08695

    work page internal anchor Pith review Pith/arXiv arXiv

  56. [56]

    Barr et al.,Salt: Multimodal Multitask Machine Learning for High Energy Physics,J

    J. Barr et al.,Salt: Multimodal Multitask Machine Learning for High Energy Physics,J. Open Source Softw.10(2025) 7217

    work page 2025

  57. [57]

    SALT documentation

    “SALT documentation.”https://ftag-salt.docs.cern.ch/, 2025

    work page 2025

  58. [58]

    Sigmoid-Weighted Linear Units for Neural Network Function Approximation in Reinforcement Learning

    S. Elfwing, E. Uchibe and K. Doya,Sigmoid-Weighted Linear Units for Neural Network Function Approximation in Reinforcement Learning,1702.03118

    work page internal anchor Pith review Pith/arXiv arXiv

  59. [59]

    On Layer Normalization in the Transformer Architecture

    R. Xiong, Y. Yang, D. He, K. Zheng, S. Zheng, C. Xing et al.,On Layer Normalization in the Transformer Architecture,2002.04745

    work page internal anchor Pith review Pith/arXiv arXiv 2002

  60. [60]

    Y. Li, D. Tarlow, M. Brockschmidt and R. Zemel,Gated Graph Sequence Neural Networks, 1511.05493

    work page internal anchor Pith review Pith/arXiv arXiv

  61. [61]

    Statistical Properties of the log-cosh Loss Function Used in Machine Learning

    R.A. Saleh and A.K.M.E. Saleh,Statistical Properties of the log-cosh Loss Function Used in Machine Learning,2208.04564. – 25 –

    work page internal anchor Pith review Pith/arXiv arXiv

  62. [62]

    Decoupled Weight Decay Regularization

    I. Loshchilov and F. Hutter,Decoupled Weight Decay Regularization,1711.05101

    work page internal anchor Pith review Pith/arXiv arXiv

  63. [63]

    C. Sun, J. Ngadiuba, M. Pierini and M. Spiropulu,Fast Jet Tagging with MLP-Mixers on FPGAs,Mach. Learn. Sci. Tech.6(2025) 035025 [2503.03103]. – 26 –

    work page arXiv 2025