Pith. sign in

REVIEW 3 major objections 5 minor 34 references

Uncertainty Quantification and Propagation for ACORN, a geometric deep learning tracking pipeline for HEP experiments

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper develops a Monte-Carlo-dropout uncertainty quantification for the ACORN GNN tracking pipeline and finds that, for the TrackML dataset, total uncertainty is dominated by aleatoric (irreducible) noise once 1,400 training events are

desk verdict Honest, useful UQ study of a GNN tracking pipeline; the empirical core is solid, but the 'full potential' claim overreaches and the abstract needs to carry the paper's own caveat about the entropy decomposition. read the letter →

arxiv 2508.16518 v1 pith:3JBKBJO6 submitted 2025-08-22 hep-ex physics.data-an

Lukas Péron , Paolo Calafiura , Xiangyang Ju , Jay Chan This is my paper · ORCID
classification hep-exphysics.data-an
keywords uncertaintyquantificationMonteCarloDropoutgraphneuralnetworksparticletrackingTrackMLaleatoricvsepistemicpropagationscorecalibration
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

The paper develops a way to measure uncertainty step-by-step through a multistage deep-learning tracking pipeline, and applies it to ACORN, a graph-neural-network particle tracker on the TrackML dataset. Using Monte Carlo Dropout to sample many weight configurations, it measures how much of the uncertainty in an edge's score is aleatoric (irreducible data noise) versus epistemic (reducible with more data or a better model). The central finding is that with 1,400 training events, the total uncertainty is already dominated by aleatoric uncertainty, meaning more training data would not substantially improve the model's confidence; and the final track-building efficiency (about 94%) is insensitive both to upstream model uncertainty (spread about 0.05%) and to calibrating the GNN's scores. A sympathetic reader should care because, if correct, it tells the experiment community where to invest next: the data side and the physics, not extra model capacity, and it provides a template for auditing other multi-stage reconstruction pipelines.

What carries the argument

Monte Carlo Dropout (MCD): during inference, dropout layers are activated T=100 times to sample the model's predictive distribution; the paper then summarizes each edge score with its empirical mean and standard deviation, splits total uncertainty (Shannon entropy of the mean score) into epistemic (mutual information) and aleatoric (conditional entropy), and uses the law of total variance to separate each stage's intrinsic uncertainty from what propagates downstream. The Filter-induced 'topological uncertainty' is handled by counting how often an edge survives the score cut across stochastic passes, so its effect on the GNN can be added up in histograms.

What would settle it

Train an ensemble of, say, 10 ACORN GNNs with different seeds on the same 1,400 events and compare each edge's ensemble variance to the MCD mutual information I[sn]; if the two diverge widely, the entropy-based split is not tracking the reducible uncertainty and the aleatoric-dominance conclusion collapses. Alternatively, increasing model capacity and re-running MCD should leave epistemic uncertainty nearly unchanged if the paper's sufficiency claim is right.

Watch

Extended reading notes

Core claim

The paper claims that for the ACORN pipeline, evaluated on TrackML, the GNN's combined uncertainty is roughly half intrinsic and half propagated from the upstream Filter, and that the total uncertainty is dominated by the aleatoric component at all training-set sizes studied (100 to 1,400 events). It further claims that ACORN is robust: the track efficiency distribution across stochastic forward passes has a spread of about 0.04 percentage points, and recalibrating edge scores with splines changes the average efficiency from 94.4% to 94.5%, not a statistically significant difference. The score posteriors are non-Gaussian, so entropy-based summaries are needed, and the results are offered as

Load-bearing premise

The entire aleatoric-dominance conclusion assumes that Shannon entropy and mutual information decompose cleanly into aleatoric and epistemic uncertainty, an additivity the paper assumes and that the cited literature questions; if that additivity fails, the claim that the model was trained to its full potential no longer follows.

Editorial extensions

If this is right

  • If aleatoric uncertainty dominates, adding more training data beyond 1,400 TrackML events is unlikely to meaningfully improve GNN confidence; improvements must come from data quality or model architecture changes.
  • The 0.05% spread in track-building efficiency implies that upstream Filter/GNN uncertainties are small enough that downstream physics analyses can treat the tracking efficiency as essentially fixed.
  • Calibration of GNN scores changes the optimal CC&Walk cut values but not the final efficiency, so readouts of raw scores as probabilities are not critical for track building performance.
  • The non-Gaussianity of MCD score posteriors motivates reporting entropy/MI-based uncertainties rather than standard deviations alone.

Reading between the lines

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

  • The same UQ&UP recipe could be applied to ACORN's metric-learning graph construction stage if real-valued adjacency matrices are introduced, an extension the paper itself flags as needing manifold-learning tools.
  • If the contested additivity assumption fails on other datasets, the qualitative conclusion (efficiency robust to dropout sampling) is likely to survive even though the aleatoric/epistemic split would need re-derivation with a different estimator.
  • A natural testable extension is to run the pipeline on HL-LHC simulation with pile-up ~200: if epistemic uncertainty grows sharply out-of-distribution, then the 'trained to full potential' claim is specific to TrackML-like conditions.
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

3 major / 5 minor

Summary. The paper develops an uncertainty quantification and propagation (UQ&P) methodology for the ACORN geometric deep learning particle tracking pipeline and applies it to the TrackML dataset. For each edge in the graph, MC Dropout with T=100 samples is used to estimate the score mean, standard deviation, and information-theoretic total, aleatoric, and epistemic uncertainties. The stochastic output of the Filter stage is propagated to the GNN by applying the score cut s_cut=0.05 over 100 stochastic graphs, and the intrinsic and propagated uncertainties are combined by histogram summation. The same stochastic procedure is propagated to the rule-based CC&Walk track builder, yielding a track efficiency distribution with a standard deviation of about 0.04%. The paper reports that GNN uncertainty is low for most edges, peaks near score 0.5, is dominated by aleatoric uncertainty, decreases with training-set size, is roughly independent of eta, increases with pT, is non-Gaussian, and is insensitive to dropout rates below 30%. Finally, calibrating the GNN edge scores does not significantly change the track reconstruction efficiency. The headline conclusions are that the ACORN pipeline was trained to its full potential and is robust to model miscalibration.

Significance. If the conclusions held, this would be a valuable demonstration that a UQ-aware characterization of a modular GNN tracking pipeline is feasible and that upstream Filter uncertainty has only a small effect on track building. The paper is honest about the contested nature of the entropy-based decomposition and ships its code on an open benchmark, which are concrete strengths. However, the central quantitative claims (aleatoric dominance, full training potential, equal split of intrinsic/propagated uncertainty) rest on a decomposition that the paper itself flags as questionable, on an arbitrary histogram-summing procedure, and on a learning curve that shows no saturation. The empirical study is a useful contribution, but the strongest claims require either additional analysis or substantial softening.

major comments (3)
  1. [Section 5.2, Eqs. (5)-(7)] The decomposition of total uncertainty into aleatoric and epistemic components via binary entropy and mutual information is not a measurement of irreducible data noise versus reducible model uncertainty for continuous edge scores. The quantity H[<s_n>] is the entropy of a Bernoulli variable with probability <s_n>, not the entropy of the predictive distribution of the score. Reference [33], cited by the authors, challenges exactly this interpretation, and the paper concedes that the results are 'indications rather than absolute truths.' Since the abstract's claim that ACORN was 'trained to its full potential' is based on this decomposition, the claim is not established. The authors should either validate the decomposition for this setting or qualify the abstract and conclusion to match the caveat in Section 5.2.
  2. [Section 6.1, Figure 15] The statement that aleatoric uncertainty dominates 'for all training dataset sizes' and that the model was trained 'to its full potential' is not supported by the data shown. The combined uncertainty is still decreasing at the largest training set of 1,400 events; no plateau, asymptote fit, or dominance threshold is reported. A decreasing total uncertainty implies that a reducible component remains. Even granting the entropy decomposition, Figure 15 does not demonstrate that additional training data would not reduce the total uncertainty further. Please add a saturation analysis or temper the conclusion accordingly.
  3. [Section 5.3, paragraph beginning 'Finally, to be able to add up...'] The method of grouping intrinsic and propagated uncertainties separately into 100-bin histograms and summing the histograms does not yield a valid per-edge combined uncertainty. The statements that the combined uncertainty is 'roughly equally distributed between the intrinsic and propagated uncertainties' and the combined distributions in Figures 10, 13, and 14 depend on this histogram summation. If per-edge values are available, the combined uncertainty should be computed per edge before any histogramming; otherwise the arbitrariness of the binning can drive the conclusion. Sensitivity to the choice of s_cut=0.05 should also be reported.
minor comments (5)
  1. [Section 5.2 and Section 6] Notation is inconsistent: the paper defines H[<s_n>] and I[s_n], but later refers to H[s_n] as the total uncertainty. Please use a single symbol, e.g., H_total[<s_n>], and distinguish it from the per-sample entropy E[H[s_n]].
  2. [Figure 15] The points in Figure 15 have no error bars or repeated-run uncertainty. With one measurement per training-set size, the decreasing trend is not quantitatively established.
  3. [Section 5.3 and Figures 9-14] The propagated-uncertainty procedure uses a fixed cut s_cut=0.05. Since the Filter's score distribution is concentrated near 0 and 1, the number of edges that pass the cut and the resulting propagated uncertainty are likely sensitive to this threshold. A brief ablation or explicit discussion would help.
  4. [Section 6.3 and Appendix A] Typographical issues: 'Shanon's entropy' should be 'Shannon's entropy'; 'non gaussianity' should be 'non-Gaussianity'; in Section 5.3, 'sGNN cut = 0 .05' contains a spurious space.
  5. [Abstract and Section 6.4] The phrase 'does not suffer from the miscalibration of the GNN model' is stronger than the evidence: the GNN scores are strongly miscalibrated (Figure 28), and the pipeline is robust only after the track-building score cuts are re-tuned. This nuance should appear in the abstract.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the UQ measurements are empirical and self-contained, with the entropy decomposition explicitly flagged as an interpretation rather than a derived truth.

full rationale

The paper's central claims are empirical measurements on the TrackML dataset using Monte Carlo Dropout. The uncertainty quantification in Section 5.2 defines total uncertainty as H[<s_n>], epistemic uncertainty as I[s_n] = H[<s_n>] - E[H[s_n]], and aleatoric uncertainty as the residual conditional entropy E[H[s_n]]. These are mathematical identities given the definitions, not circular reductions. The dominance of aleatoric uncertainty is then a direct measurement of these quantities, not a quantity fitted to force the conclusion. The step from dominance to 'sufficient data to train the ACORN model to its full potential' relies on the interpretation that the conditional-entropy term is irreducible, but the paper explicitly acknowledges this interpretation is disputed by Ref. [33] and states the results 'should be considered as indications rather than absolute truths.' That is an honest caveat about the validity of an imported measure, not a circular argument. The pipeline itself (ACORN) is prior work cited from the same collaboration, but it is used as the object of study, not as evidence for the UQ conclusions. No parameter is fitted to a subset of data and then called a prediction; the track-efficiency distribution is built from repeated stochastic forward passes and compared to external truth labels. Overall, the reasoning is self-contained and empirically grounded, so the circularity score is 0.

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

The paper's central claims rest on the MCD approximation, the questionable additive entropy decomposition, the representativeness of TrackML, and a few hand-picked thresholds (Filter cut 0.05, dropout 0.1). No new physical entities are introduced.

free parameters (5)
  • s_cut (Filter-to-GNN edge pruning threshold) = 0.05
    Chosen threshold for propagating Filter uncertainty to GNN; affects the measured balance between intrinsic and propagated GNN uncertainty.
  • Dropout rate for MCD = 0.1 (training and nominal inference)
    Chosen classical value; paper tests 0.05-0.95 and finds stability below 0.3, so central claim is robust to this choice.
  • CC&Walk score cuts (connected components, minimal branching, additive branching) = uncalibrated: 0.01, 0.1, 0.6; calibrated: 0.5, 0.9, 1.0
    Tuned per case using the calibration curves; necessary for the efficiency comparison.
  • Number of MCD samples T = 100
    Standard choice; no sensitivity analysis provided.
  • Histogram bin count for combining uncertainties = 100
    Chosen binning for summing intrinsic and propagated uncertainty histograms; affects the visual and quantitative 'equally distributed' claim.
assumptions (4)
  • domain assumption MCD with dropout at inference approximates the posterior distribution of the model weights
    The entire UQ methodology relies on this Bayesian approximation (Gal & Ghahramani, Ref [11]). Section 5.2.
  • ad hoc to paper Total uncertainty decomposes additively into aleatoric and epistemic components via entropy and mutual information
    The paper states this assumption explicitly and cites Ref [33] which questions the appropriateness of these measures. Section 5.2.
  • domain assumption The TrackML dataset is representative of HL-LHC ITk tracking conditions
    All experiments are on TrackML; the conclusions about ACORN robustness are framed in the HL-LHC context. Section 5.1.
  • domain assumption The Filter and GNN models are fully trained to convergence
    Additivity and epistemic interpretation assume fully trained models; paper states 'we work only with fully trained model'. Section 5.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Uncertainty Quantification and Propagation for ACORN, a geometric deep learning tracking pipeline for HEP experiments." pith.science (2026). https://pith.science/paper/3JBKBJO6

@misc{pith2026250816518,
  author       = {Pith},
  title        = {Pith review of: Uncertainty Quantification and Propagation for ACORN, a geometric deep learning tracking pipeline for HEP experiments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3JBKBJO6}},
  note         = {Machine review of arXiv:2508.16518}
}
read the original abstract

We have developed an Uncertainty Quantification process for multistep pipelines and applied it to the ACORN particle tracking pipeline. All our experiments are made using the TrackML open dataset. Using the Monte Carlo Dropout method, we measure the data and model uncertainties of the pipeline steps, study how they propagate down the pipeline, and how they are impacted by the training dataset's size, the input data's geometry and physical properties. We will show that for our case study, as the training dataset grows, the overall uncertainty becomes dominated by aleatoric uncertainty, indicating that we had sufficient data to train the ACORN model we chose to its full potential. We show that the ACORN pipeline yields high confidence in the track reconstruction and does not suffer from the miscalibration of the GNN model.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 30 canonical work pages

  1. [33]

    Wimmer, Y

    L. Wimmer, Y. Sale, P. Hofman, B. Bis- chl, and E. Hüllermeier. Quantifying aleatoric and epistemic uncertainty in ma- chine learning: Are conditional entropy and mutual information appropriate mea- sures?, 2023

    work page 2023

  2. [1]

    Observation of a new particle in the search for the standard model higgs boson with the at- las detector at the lhc.Physics Letters B, 716(1):1–29, 2012

    The ATLAS collaboration. Observation of a new particle in the search for the standard model higgs boson with the at- las detector at the lhc.Physics Letters B, 716(1):1–29, 2012

    work page 2012

  3. [2]

    Observation of a new boson at a mass of 125 gev with the cms experiment at the lhc.Physics Letters B, 716(1):30–61, 2012

    CMS Collaboration. Observation of a new boson at a mass of 125 gev with the cms experiment at the lhc.Physics Letters B, 716(1):30–61, 2012

    work page 2012

  4. [3]

    Laura Gonella. The atlas itk detector sys- tem for the phase-ii lhc upgrade.Nuclear Instruments and Methods in Physics Re- search Section A: Accelerators, Spectrom- eters, Detectors and Associated Equip- ment, 1045:167597, 2023

    work page 2023

  5. [4]

    Expected tracking performance of the atlas inner trackeratthehigh-luminositylhc

    The ATLAS collaboration. Expected tracking performance of the atlas inner trackeratthehigh-luminositylhc. Journal of Instrumentation, 20(02):P02018, feb 2025

    work page 2025

  6. [5]

    Bronstein, Joan Bruna, Taco Cohen, and Petar Veličković

    Michael M. Bronstein, Joan Bruna, Taco Cohen, and Petar Veličković. Geomet- ric deep learning: Grids, groups, graphs, geodesics, and gauges, 2021

    work page 2021

  7. [6]

    Performance of a Geometric Deep Learning Pipeline for HL-LHC Particle Tracking

    Xiangyang Ju, Daniel Murnane, Paolo Calafiura, Nicholas Choma, Sean Con- lon, Steve Farrell, Yaoyuan Xu, Maria Spiropulu, Jean-Roch Vlimant, Adam Au- risano, Jeremy Hewes, Giuseppe Cerati, Lindsey Gray, Thomas Klijnsma, Jim Kowalkowski, Markus Atkinson, Mark Neubauer, Gage DeZoort, Savannah Thais, Aditi Chauhan, Alex Schuy, Shih- Chieh Hsu, and Alex Ballo...

    arXiv 2021

  8. [7]

    Towards a realistic track reconstruction algorithm based on graph neural networks for the hl-lhc.EPJ Web Conf., 251:03047, 2021

    Biscarat, Catherine, Caillou, Sylvain, Rougier, Charline, Stark, Jan, and Zahreddine, Jad. Towards a realistic track reconstruction algorithm based on graph neural networks for the hl-lhc.EPJ Web Conf., 251:03047, 2021. 15

    work page 2021

Show all 34 references
  1. [8]

    ATLAS ITk Track Reconstruction with a GNN-based pipeline

    Sylvain Caillou, Paolo Calafiura, Steven Andrew Farrell, Xiangyang Ju, Daniel Thomas Murnane, Charline Rougier, Jan Stark, and Alexis Vallier. ATLAS ITk Track Reconstruction with a GNN-based pipeline. Technical report, CERN, Geneva, 2022

    2022

  2. [9]

    Acorn - a charged object reconstruc- tion network.https://gitlab.cern.ch/ gnn4itkteam/acorn/

    Markus Julian Atkinson, Jackson Burzyn- ski, Pierfrancesco Butti, Jared Burleson, Sylvain Caillou, Paolo Calafiura, Jay Chan, Christophe Collard, Sebastian Dittmeier, Steven Andrew Farrell, Ben- jamin Huth, Xiangyang Ju, Alina Lazar, Ryan Liu, Tuan Minh Pham, Daniel Murnane, M...

  3. [10]

    ATLAS Soft- ware and Computing HL-LHC Roadmap

    The ATLAS collaboration. ATLAS Soft- ware and Computing HL-LHC Roadmap. Technical report, CERN, Geneva, 2022

    2022

  4. [11]

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning, 2016

    Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning, 2016

    2016

  5. [12]

    Trackml particle tracking challenge

    Anaderi, Andreas Salzburger, Ceci- leGermain, David Rousseau, Guillaume Charpiat, Heather Gray, inversion, Is- abelle, JR, Laurent Basara, Maggie, Mikhail Hushchyn, Moritz Kiehn, Paolo Calafiura, Sorme, and Steve Farrell. Trackml particle tracking challenge. https://kaggle.com...

  6. [13]

    Axiomatic attribution for deep networks, 2017

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks, 2017

    2017

  7. [14]

    Fatal crash between a car operating with auto- mated control systems and a tractor- semitrailer truck

    Kristin Poland, Mary Pat McKay, Deb Bruce, and Ensar Becic. Fatal crash between a car operating with auto- mated control systems and a tractor- semitrailer truck. Traffic Injury Preven- tion, 19(sup2):S153–S156, 2018. PMID: 30841795

    2018

  8. [15]

    Bayesian neural networks, 2020

    Tom Charnock, Laurence Perreault- Levasseur, and François Lanusse. Bayesian neural networks, 2020

    2020

  9. [16]

    Maximizing overall di- versity for improved uncertainty estimates in deep ensembles, 2020

    Siddhartha Jain, Ge Liu, Jonas Mueller, and David Gifford. Maximizing overall di- versity for improved uncertainty estimates in deep ensembles, 2020

    2020

  10. [17]

    Evidential deep learning to quantify classification uncertainty, 2018

    Murat Sensoy, Lance Kaplan, and Melih Kandemir. Evidential deep learning to quantify classification uncertainty, 2018

    2018

  11. [18]

    Marius Zöllner

    Christian Hubschneider, Robin Hut- macher, and J. Marius Zöllner. Calibrat- ing uncertainty models for steering an- gle estimation. In 2019 IEEE Intelli- gent Transportation Systems Conference (ITSC), pages 1511–1518, 2019

    2019

  12. [19]

    Andrew Y. K. Foong, David R. Burt, Yingzhen Li, and Richard E. Turner. On the expressiveness of approximate infer- ence in bayesian neural networks, 2020

    2020

  13. [20]

    An imple- mentation of neural simulation-based in- ference for parameter estimation in at- las

    The ATLAS collaboration. An imple- mentation of neural simulation-based in- ference for parameter estimation in at- las. Reports on Progress in Physics , 88(6):067801, May 2025

    2025

  14. [21]

    Deep neural network un- certainty quantification for lartpc recon- struction, 2023

    Dae Heun Koh, Aashwin Mishra, and Kazuhiro Terao. Deep neural network un- certainty quantification for lartpc recon- struction, 2023

    2023

  15. [22]

    Evidential deep learning for uncertainty quantification and out-of- distribution detection in jet identification using deep neural networks

    Ayush Khot, Xiwei Wang, Avik Roy, Volodymyr Kindratenko, and Mark Neubauer. Evidential deep learning for uncertainty quantification and out-of- distribution detection in jet identification using deep neural networks. Machine Learning: Science and Technology, 2025

    2025

  16. [23]

    Chained machine learning model for predicting load capac- ity and ductility of steel fiber–reinforced 16 concrete beams

    Torkan Shafighfard, Farzin Kazemi, Fara- marz Bagherzadeh, Magdalena Mieloszyk, and Doo-Yeol Yoo. Chained machine learning model for predicting load capac- ity and ductility of steel fiber–reinforced 16 concrete beams. Computer-Aided Civil and Infrastructure Engineering , 39(2...

    2024

  17. [24]

    Multi- phase flow rate prediction using chained multi-output regression models

    Md Ferdous Wahid, Reza Tafreshi, Zurwa Khan, and Albertus Retnanto. Multi- phase flow rate prediction using chained multi-output regression models. Geoen- ergy Science and Engineering, 231:212403, 2023

    2023

  18. [25]

    Com- paring a composite model versus chained modelstolocateanearestvisualobject

    Antoine Le Borgne, Xavier Marjou, Fanny Parzysz, and Tayeb Lemlouma. Com- paring a composite model versus chained modelstolocateanearestvisualobject. In 2023 IEEE/ACIS 8th International Con- ference on Big Data, Cloud Computing, and Data Science (BCD), pages 231–236, 2023

    2023

  19. [26]

    Fast dropout training

    Sida Wang and Christopher Manning. Fast dropout training. In Sanjoy Das- gupta and David McAllester, editors,Pro- ceedings of the 30th International Con- ference on Machine Learning, volume 28 of Proceedings of Machine Learning Re- search, pages 118–126, Atlanta, Georgia, USA,...

    2013

  20. [27]

    Un- certainty quantification via stable distribution propagation, 2024

    Felix Petersen, Aashwin Mishra, Hilde Kuehne, Christian Borgelt, Oliver Deussen, and Mikhail Yurochkin. Un- certainty quantification via stable distribution propagation, 2024

    2024

  21. [28]

    Uncertainty propagation within chained models for machine learning re- construction of neutrino-lar interactions, 2025

    Daniel Douglas, Aashwin Mishra, Daniel Ratner, Felix Petersen, and Kazuhiro Terao. Uncertainty propagation within chained models for machine learning re- construction of neutrino-lar interactions, 2025

    2025

  22. [29]

    L. Kong, J. Sun, and C. Zhang. Sde-net: Equipping deep neural networks with un- certainty estimates, 2020

    2020

  23. [30]

    Acommontrackingsoftwareproject

    Xiaocong Ai, Corentin Allaire, Noemi Calace, Angéla Czirkos, Markus Elsing, Irina Ene, Ralf Farkas, Louis-Guillaume Gagnon, Rocky Garg, Paul Gessinger, etal. Acommontrackingsoftwareproject. Computing and Software for Big Science, 6(1):8, 2022

    2022

  24. [31]

    The ATLAS collaboration. Athena. https://doi.org/10.5281/zenodo. 2641997, April 2019

    2019 doi

  25. [32]

    Uncertainty in Deep Learn- ing

    Yarin Gal. Uncertainty in Deep Learn- ing. PhD thesis, University of Cambridge, 2016

    2016

  26. [34]

    Nixon, M

    J. Nixon, M. Dusenberry, G. Jerfel, T. Nguyen, J. Liu, L. Zhang, and D. Tran. Measuring calibration in deep learning, 2020. A Filter UQ procedure Figure 32: Filter ROC curve. 17 Figure 33: Filter efficiency vsη. Figure 34: Mean edge score⟨sFilter n ⟩. Figure 35: Filter uncerta...

    2020

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.