Pith. sign in

REVIEW 4 major objections 5 minor 37 references

Applicability Evaluation of Selected xAI Methods for Machine Learning Algorithms for Signal Parameters Extraction

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read Occlusion sensitivity reveals what a pulse-reconstruction autoencoder relies on, and points to a roughly 30-nanosecond limit for separating two pulses.

desk verdict A useful qualitative xAI heuristic for pulse-reconstruction autoencoders, but the 30 ns separation estimate is an unsupported extrapolation. read the letter →

arxiv 2504.15670 v1 pith:56J7LGBR submitted 2025-04-22 physics.comp-ph hep-exphysics.ins-det

classification physics.comp-phhep-exphysics.ins-det
keywords explainableAIxocclusionsensitivityconvolutionalautoencoderpulsereconstructionscintillatingcrystalwaveformssignalparameterextractiondouble-pulseseparation
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 asks which explainable-AI method actually helps physicists understand what a pulse-reconstruction autoencoder has learned, and whether that understanding can guide detector design. It applies four attribution methods - Integrated Gradients, Vanilla Saliency, SmoothGrad, and Occlusion Sensitivity - to three variants of a convolutional autoencoder trained on simulated scintillator-crystal waveforms. It reports that Occlusion Sensitivity gives the clearest and most usable picture: unmasking a recorded pulse bit by bit produces a sharp minimum in the reconstruction loss, and the position of that minimum aligns with the size of the model's final convolutional filter. From that alignment the paper estimates the minimum time separation at which two pulses can still be told apart, about 30 nanoseconds. The intended payoff is that xAI can turn a black-box pulse finder into a diagnostic tool for architecture choices and detector timing requirements.

What carries the argument

The central object is a modified convolutional autoencoder with an encoder of three 1D convolution layers and a decoder of three transposed-convolution layers, where the final filter has kernel size 18, with variants using kernel size 14 and linear rather than ReLU output. The carrying mechanism is Occlusion Sensitivity in a progressive form: one unmasked sample at a time after the pulse rise, while plotting the reconstruction total loss versus the number of unmasked values. The position of the loss minimum is interpreted as the minimal waveform information required for detection, and its coincidence with the last filter's kernel size is what links architecture to the inferred 30-ns two-pulse separation limit.

What would settle it

Measure the actual sampling rate of the simulated waveforms (or re-run the unmasking experiment at a stated 1 sample/ns); if 18 samples do not correspond to roughly 20 ns of rise time plus kernel width, the 30 ns separation estimate changes. Alternatively, test M(18, ReLu) on pairs of pulses separated by less than 30 ns: if it can reconstruct both pulses, the estimated minimum separation is not a true limit.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that Occlusion Sensitivity is the xAI method that yields actionable understanding of the modified autoencoder's pulse reconstruction. When a single recorded pulse is progressively unmasked, the total reconstruction loss drops to a sharp minimum and then plateaus; for the model with an 18-sample last-filter kernel and ReLU output, the minimum occurs after 18 samples have been unmasked, matching the kernel size. The paper reads this as the minimal waveform window the model needs for reliable detection and converts it, together with the 10-ns pulse rise time, into an estimated minimum temporal separation of about 30 ns between two distinguishable signals. It also reports that the ReLU output layer confines model activity to pulse arrival positions, while linear output leaves activity across the whole noise-dominated waveform.

Load-bearing premise

The central quantitative claim assumes that the lowest total-loss point in the progressive-unmasking curve marks the smallest waveform window the network needs, and that the number of samples at that point can be converted into nanoseconds even though the sampling rate of the 1024-sample, 1024-ns waveforms is never stated in the paper.

Editorial extensions

If this is right

  • Occlusion-based unmasking can identify the exact waveform region (rise and peak) that drives a pulse-reconstruction network, so future models can focus training and architecture on that region.
  • The coincidence between loss-minimum position and the last filter's kernel size gives a way to estimate the model's effective receptive field from explanation output rather than from parameter counting.
  • The inferred roughly 30-ns minimal two-pulse separation would serve as a design target for calorimeter readout and pile-up rejection requirements.
  • Models with shallower minima reaching their minimum earlier may be better suited for separating closely spaced pulses, a hypothesis the paper suggests for follow-up.
  • The last-layer activation function is singled out as a control on whether the model localizes pulses or produces noisy all-event output, which matters for pulse counting and arrival-time determination.

Reading between the lines

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

  • A direct test of the 30 ns number is to re-run the unmasking experiment at an explicitly stated sampling rate; if the rate is not 1 sample/ns, the inferred separation should scale accordingly.
  • The kernel-size coincidence suggests a more general diagnostic: for any convolutional pulse finder, the progressive-occlusion minimum could be used to measure receptive field and minimum distinguishable signal spacing without retraining.
  • Comparing the three models suggests activation function and kernel size are independent tuning knobs; a follow-up could sweep kernel sizes at fixed activation to see whether the loss-minimum position tracks kernel size monotonically.
  • The same progressive-unmasking procedure could be applied directly to two-pulse events, masking one pulse and measuring how many samples are needed to detect the other, which would give a per-event double-pulse resolution estimate.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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. The manuscript evaluates four explainable-AI techniques—Integrated Gradients, Vanilla Saliency, SmoothGrad, and Occlusion Sensitivity—on modified convolutional autoencoders that reconstruct simulated scintillating-crystal pulses. Three model variants are considered, differing in the last-filter kernel size (18 vs 14) and the output activation (linear vs ReLU): M(18, lin), M(14, lin), and M(18, ReLU). After presenting layer-output plots and saliency maps, the authors focus on occlusion sensitivity, including a progressive-unmasking experiment in which a single pulse is gradually revealed. They report that the signal rise and peak are the most loss-relevant regions, that M(18, ReLU) reaches the deepest loss minimum after 18 unmasked values, and that this minimum coincides with the last-filter kernel size. They further estimate a minimum resolvable separation of 30 ns between two pulses and conclude that occlusion sensitivity is the most informative xAI method for this application.

Significance. The qualitative observation that occlusion-based attribution highlights the pulse rise and peak, rather than the long decay tail, is a useful demonstration of xAI in a HEP waveform-reconstruction setting, and the progressive-unmasking diagnostic is a promising idea for probing receptive fields. The paper also gives a concrete three-way architecture comparison and states the SmoothGrad hyperparameters, which helps reproducibility of that part. If the quantitative claims were backed by a defined loss, error bars, and a direct two-pulse test, the paper would make a meaningful contribution to applied xAI in calorimetry. As it stands, the contribution is a promising proof-of-concept whose headline numbers are not yet supported.

major comments (4)
  1. [Section 3.4, Figs. 5-6] The 'total loss' plotted in the occlusion-sensitivity figures is never defined. The manuscript does not state which loss function is used, whether the curves are for a single event or averaged over the test set, or how the mask is applied in the progressive-unmasking variant. Without this information the minima in Fig. 6 cannot be interpreted as the 'minimum information required' for detection, and the comparison among models is not reproducible. Please specify the loss, the aggregation, and the event-to-event spread (e.g., mean plus/minus standard deviation or percentiles over many events).
  2. [Section 4] The conversion of 18 unmasked values into 30 ns is not justified. The waveforms are described as 1024 ns long, but the sampling rate is never given. If the digitization is one sample per nanosecond, that must be stated explicitly, and the relationship '18 samples + 10 ns rise approximately equals 30 ns' needs an explicit derivation. As written, the headline quantitative result is not reproducible.
  3. [Section 4 vs. Section 2] The estimate of 30 ns as the minimum separation between two distinguishable signals is not supported by any two-pulse experiment on the three models studied. The only empirical anchor cited, ref. [34], reports that pulses more than 10 ns apart can be separated for an earlier model, which is also in tension with the new 30 ns estimate. The single-pulse unmasking minimum may be related to the receptive field, but no argument is given that it equals the two-pulse resolution limit. A direct double-pulse test, or an explicit validated mapping, is required before the 30 ns claim can stand.
  4. [Section 4] The claimed coincidence between the 18-unmasked-value minimum and the last-filter kernel size is under-identified. M(18, ReLU) differs from M(14, lin) in both kernel size and activation function, so the minimum position cannot be attributed to kernel size alone. The comparison of M(18, lin) with M(14, lin) partially isolates the kernel size, but the specific 18-value coincidence is claimed for M(18, ReLU), which is confounded by activation. A control with a single varied factor (for example, adding M(14, ReLU)) is needed to support the kernel-size interpretation.
minor comments (5)
  1. [Section 3] Integrated Gradients is listed among the tested methods, but no Integrated Gradients results are presented in Section 3; please add the results or revise the claim.
  2. [Figure 5 caption] The caption contains a typo ('fuction') and should be corrected to 'function'.
  3. [Eq. (1)] Equation (1) uses the notation 'M_c(x)' without defining 'M_c' or the subscript 'c'; please define all symbols before using them.
  4. [References] References [24] and [26] are the same paper; please consolidate and remove the duplicate.
  5. [Section 4, final paragraph] The statement that models with shallower minima 'might be better suited' for separating closely spaced signals is speculative and is not tested in the manuscript; consider removing it or supporting it with a two-pulse experiment.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: xAI outputs are computed from trained models and the 18-value/kernel-size coincidence is an observed correlation, not a fitted input.

full rationale

The paper does not derive its central xAI conclusion from a parameter that was fitted to that conclusion. The models are trained on simulated scintillator pulses, and the four xAI attributions are computed post hoc from the frozen networks. The occlusion experiment (Section 3.4, Figure 6) progressively unmasks samples and plots the resulting total loss; the location of the minimum (18 unmasked values for M(18, ReLu)) is an empirical observable, not an input to any fit. The claimed coincidence with the last-layer kernel size and the inferred 30 ns separation limit are interpretive after-the-fact statements, not reductions of the output to the input. The only self-references are [33]-[35], which supply the model architecture and the previous 10 ns separation result; these citations are not used to justify the new xAI claims, and the xAI evaluation itself is self-contained with respect to the trained models. The concern that the 30 ns estimate is unsupported (missing sampling-rate conversion, no direct double-pulse test) is a correctness/validity issue, not a circularity issue, and therefore does not raise the circularity score.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

No new physical entities are postulated. The ledger's content is architectural parameters and interpretive assumptions: the central quantitative claims (kernel-size coincidence, 30 ns separation) are read off loss curves produced by hand-chosen mask sizes on a synthetic dataset, not measured against physics.

free parameters (3)
  • Occlusion mask size = 18 consecutive samples
    Chosen by hand in Section 3.4; the mask width controls the loss curves and the position of the observed minima from which the kernel-size coincidence is read off.
  • SmoothGrad noise sigma and averaging window n = sigma = half of the maximum signal value of each event; n = 10
    Chosen by hand in Section 3.3; these are standard SmoothGrad hyperparameters and do not enter the central occlusion analysis, but they determine the reported SmoothGrad saliency values.
  • Last-filter kernel sizes = 18 (M(18, lin), M(18, ReLu)); 14 (M(14, lin))
    Architecture choices from the authors' earlier work [33]; the claim that the loss minimum coincides with the kernel size (18) and the derived 30 ns separation estimate depend on these specific values.
assumptions (3)
  • domain assumption The simulated waveforms (1024 ns, up to 4 pulses, 10 ns rise, 300 ns fall, Gaussian amplitudes 200+-200 mV, 10 mV mean noise) capture the essential structure of electromagnetic calorimeter signals.
    Section 2; the entire analysis and all physics conclusions are evaluated only on this synthetic dataset, and the connection to real PADME-style signals is inherited from the authors' earlier papers [34, 35].
  • domain assumption The total loss used in occlusion sensitivity (unstated, implicitly reconstruction error against the label) is a valid proxy for signal detection quality.
    Section 3.4; the paper never states the loss function, yet all conclusions about which waveform regions matter and which model 'performs best' are drawn from the magnitude and shape of this loss.
  • ad hoc to paper The point of minimum total loss during progressive unmasking marks the minimum waveform information required for reliable detection, and the number of unmasked samples maps 1:1 to nanoseconds.
    Section 4; this interpretive step converts an internal loss curve into the physical claim of 30 ns minimum temporal separation. The sampling rate (samples per ns) is never stated, and the 10 ns separation figure from [34] in Section 2 is not reconciled with the 30 ns estimate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Applicability Evaluation of Selected xAI Methods for Machine Learning Algorithms for Signal Parameters Extraction." pith.science (2026). https://pith.science/paper/56J7LGBR

@misc{pith2026250415670,
  author       = {Pith},
  title        = {Pith review of: Applicability Evaluation of Selected xAI Methods for Machine Learning Algorithms for Signal Parameters Extraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/56J7LGBR}},
  note         = {Machine review of arXiv:2504.15670}
}
read the original abstract

Machine learning methods find growing application in the reconstruction and analysis of data in high energy physics experiments. A modified convolutional autoencoder model was employed to identify and reconstruct the pulses from scintillating crystals. The model was further investigated using four xAI methods for deeper understanding of the underlying reconstruction mechanism. The results are discussed in detail, underlining the importance of xAI for knowledge gain and further improvement of the algorithms.

Figures

Figures reproduced from arXiv: 2504.15670 by the authors.

Figure 1
Figure 1. (Up, left:) A simulated event with no pulses. The waveform contains only the gaussian noise. (Up, right:) A simulated event with two pulses, corresponding to two particles entering a crystal. (Down, left:) The output of applying the autoencoder network to the event from the upper right panel. The result follows the shape of the input data waveform while denoising it in the signal regions. (Down, right:) The result o… view at source ↗
Figure 2
Figure 2. Layers output for the three tested models for an event with two pulses created by two particles [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Vanilla Saliency results for the three modified autoencoder models. All three show rapidly [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Results from applying the SmoothGrad method to the output of the modified autoencoder [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Results from applying the Occlusion Sensitivity method to the three modified autoencoder [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Results from the modified occlusion sensitivity application to an event with one pulse. For all [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 25 canonical work pages

  1. [34]

    Dimitrova, K.; et al. (2022). Using Artificial Intelligence in the Reconstruction of Signals from the PADME Electromagnetic Calorimeter. Instruments 6, 46

  2. [1]

    Whiteson, S., & Whiteson, D. (2009). Machine learning for event selection in high-energy physics. Engineering Applications of Artificial Intelligence, Volume 22, Issue 8, 2009, Pages 1203-1217, ISSN 0952-1976, https://doi.org/10.1016/j.engappai.2009.05.004

  3. [2]

    Pang, L. (2021). Feature extraction in heavy-ion collisions using machine learning. Nuclear Physics A, 1005 , 121972, ISSN 0375-9474, https://doi.org/10.1016/j.nuclphysa.2020.121972

  4. [3]

    Karagiorgi, G., Kasieczka, G., Kravitz, S. et al. Machine learning in the search for new fundamental physics. Nat Rev Phys 4, 399–412 (2022). https://doi.org/10.1038/s42254-022-00455-1

  5. [4]

    Guest, D., et al. (2018). Deep learning and its application to LHC physics. Annual Review of Nuclear and Particle Science, 68 , 161–181. https://doi.org/10.1146/annurev-nucl-101917-021019

  6. [5]

    Kai Zhou, Lingxiao Wang, Long-Gang Pang, Shuzhe Shi, Exploring QCD matter in extreme conditions with Machine Learning, Progress in Particle and Nuclear Physics, Volume 135, 2024, 104084, ISSN 0146-6410, https://doi.org/10.1016/j.ppnp.2023.104084

  7. [6]

    Pang, Long-Gang, Studying high-energy nuclear physics with machine learning, International Journal of Modern Physics E, 33, 06, 2430009, 2024, https://doi.org/10.1142/S0218301324300091

  8. [7]

    He, WB., Ma, YG., Pang, LG. et al. High-energy nuclear physics meets machine learning. NUCL SCI TECH 34, 88 (2023). https://doi.org/10.1007/s41365-023-01233-z

Show all 37 references
  1. [8]

    Rasiklal Yadav, B. (2024). The Ethics of Understanding: Exploring Moral Implications of Explainable AI. International Journal of Science and Research (IJSR)

  2. [9]

    Zodage, P., Harianawala, H., Shaikh, H., & Kharodia, A. (2024). Explainable AI (xAI): History, Basic Ideas and Methods. International Journal of Advanced Research in Science, Communication and Technology

  3. [10]

    Kamath, U., Liu, J. (2021). Introduction to Interpretability and Explainability. In: Explain- able Artificial Intelligence: An Introduction to Interpretable Machine Learning. Springer, Cham. https://doi.org/10.1007/978-3-030-83356-5 1

  4. [11]

    Kumawat, K. (2024). Explainable AI: Interpretable Models for Transparent Decision-Making. inter- national journal of food and nutritional sciences.,doi:10.48047/ijfans/09/03/30

  5. [12]

    Sewada, R., Jangid, A., Kumar, P., Mishra, N. (2023). Explainable Artificial Intelligence (xAI). international journal of food and nutritional sciences., doi:10.48047/ijfans/v12/i1/271

  6. [13]

    (2023).The emergence and need for explainable AI.Advances in Engineering Innovation,3,1-4., doi:10.54254/2977-3903/3/2023023

    Chia,H.L.B. (2023).The emergence and need for explainable AI.Advances in Engineering Innovation,3,1-4., doi:10.54254/2977-3903/3/2023023

  7. [14]

    Tiwari, R. (2023). Explainable AI (xAI) and its Applications in Building Trust and Understanding in AI Decision Making. INTERANTIONAL JOURNAL OF SCIENTIFIC RESEARCH IN ENGI- NEERING AND MANAGEMENT.,doi:10.55041/ijsrem17592

  8. [15]

    Adhikari, T. (2023). Towards Explainable AI: Interpretable Models and Feature Attribution. SSRN Electronic Journal., doi:10.2139/ssrn.4376176

  9. [16]

    Smilkov, D., Thorat, N., Kim, B., Vi´ egas, F.B., & Wattenberg, M. (2017). SmoothGrad: removing noise by adding noise. ArXiv, abs/1706.03825

  10. [17]

    Valois, P., Niinuma, K., & Fukui, K. (2023). Occlusion Sensitivity Analysis with Augmentation Subspace Perturbation in Deep Feature Space. 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 4817-4826

  11. [18]

    Mundhenk, T.N., Chen, B.Y., & Friedland, G. (2019). Efficient Saliency Maps for Explainable AI. ArXiv, abs/1911.11293

  12. [19]

    Seo, J., Choe, J., Koo, J., Jeon, S., Kim, B., & Jeon, T. (2018). Noise-adding Methods of Saliency Map as Series of Higher Order Partial Derivative. ArXiv, abs/1806.03000

  13. [20]

    Hsu, C., Li, W. (2023). Explainable GeoAI: can saliency maps help interpret artificial intelligence’s learning process? An empirical study on natural feature detection. International Journal of Geograph- ical Information Science, 37, 963 - 987

  14. [21]

    Yeh, C., Hsieh, C., Suggala, A.S., Inouye, D.I., Ravikumar, P. (2019). On the (In)fidelity and Sensitivity for Explanations. arXiv: Learning

  15. [22]

    Konate, S., Lebrat, L., Cruz, R.S., Smith, E., Bradley, A., Fookes, C., Salvado, O. (2021). A Comparison of Saliency Methods for Deep Learning Explainability. 2021 Digital Image Computing: Techniques and Applications (DICTA), 01-08

  16. [23]

    Pezoa, R., Salinas, L., Torres, C.E., Mar ´ıa, F.S. (2023). Explainability of High Energy Physics events classification using SHAP. Journal of Physics: Conference Series, 2438. 10.1088/1742- 6596/2438/1/012082

  17. [24]

    Neubauer, M.S., Roy, A. (2022). Explainable AI for High Energy Physics. ArXiv, abs/2206.06632

  18. [25]

    Maglianella, L., Nicoletti, L., Giagu, S., Napoli, C., Scardapane, S. (2023). Convergent Approaches to AI Explainability for HEP Muonic Particles Pattern Recognition. Computing and Software for Big Science, 7, 1-18

  19. [26]

    Neubauer, M.S., Roy, A. (2022). Explainable AI for High Energy Physics. ArXiv, abs/2206.06632. 10.48550/arXiv.2206.06632

  20. [27]

    Turvill, D., Barnby, L.S., Yuan, B., Zahir, A. (2020). A Survey of Interpretability of Machine Learning in Accelerator-based High Energy Physics. 2020 IEEE/ACM International Conference on Big Data Computing, Applications and Technologies (BDCAT), 77-86

  21. [28]

    C´ ardenas-Montes, M., Santorelli, R. (2024). Neural Networks for position reconstruction in liquid argon detectors. Journal of Instrumentation, 19

  22. [29]

    Lai, Y.S., Neill, D., Plosko’n, M., Ringer, F.M. (2020). Explainable machine learning of the under- lying physics of high-energy particle collisions. Physics Letters B

  23. [30]

    Gupta, R., Bhattacharya, T., Yoon, B. (2022). AI and Theoretical Particle Physics

  24. [31]

    and Nash, R

    O’Shea, K. and Nash, R. (2015). An Introduction to Convolutional Neural Networks. arXiv:cs.NE/1511.08458

  25. [32]

    Zhang, Y. (2018). A better autoencoder for image: Convolutional autoencoder. In Proceedings of the ICONIP17-DCEC , Oct 14 – 18, 2017, Guangzhou, China

  26. [33]

    Buchakchiev, V.; et al. (2023). Pattern recognition and signal parameters extraction using machine learning methods. J. Phys.: Conf. Ser. 2668, 012001

  27. [35]

    Dimitrova, K.; et al. (2024). Machine learning assisted reconstruction of positron-on-target annihi- lation events in the PADME experiment. J. Phys.: Conf. Ser. , 2794, 012001

  28. [36]

    Sundararajan, M. (2017). Axiomatic attribution for deep networks. ICML’17: Proceedings of the 34th International Conference on Machine Learning 70, 3319 - 3328

  29. [37]

    Simonyan, K.; et al. (2014). Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps. Workshop at International Conference on Learning Representations

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.