Pith. sign in

REVIEW 4 major objections 6 minor 33 references

Continual Learning via Ensemble-Based Depth-Wise Masked Autoencoders for Data Quality Monitoring in High-Energy Physics

T0 review · 4 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a lightweight, continually refreshed ensemble of depth-wise masked autoencoders keeps anomaly-detection precision above 98.8% on CMS detector data even as the data distribution shifts between 2018 and 2022, without th

desk verdict Useful controlled study of a lightweight continual-learning DQM system, but the pristine-data recalibration assumption is load-bearing and the evaluation is narrower than the abstract implies. read the letter →

arxiv 2603.02369 v2 pith:YQPYN3LO submitted 2026-03-02 hep-ex

Dale Julson , Eric Reinhardt , Andrii Krutsylo , Resham Sohal , Guillermo Fidalgo , Sergei Gleyzer , Emanuele Usai , The CMS HCAL Collaboration This is my paper
classification hep-ex
keywords continualmachinelearninganomalydetectiondataqualitymonitoringmaskedautoencoderdepth-wiseattentiondistributionshiftCMSHCALensemble
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 tackles a practical problem in high-energy physics: machine-learning models that monitor detector data degrade as the data distribution shifts over time. It introduces DepthViT, a masked autoencoder with roughly 300,000 parameters, and combines two continual learning tricks: refreshing the error statistics (mu and sigma) on each new run's pristine data, and ensembling four DepthViT models trained on different recent runs. The paper reports that this combination holds precision above 98.8% and recall near 1.0 across 2018 and 2022 CMS hadron calorimeter data, while a static model's false-negative rate climbs to 0.75 on the shifted 2022 runs. If true, this offers a low-cost path to keeping monitoring systems reliable over long detector lifetimes and in any environment where data drifts.

What carries the argument

DepthViT (depth-wise masked autoencoder): a vision-transformer-style autoencoder whose patch embedding uses depth-wise convolution (each input channel gets its own filters, rather than summing over channels) and whose self-attention computes attention weights along the channel/depth axis instead of the spatial sequence axis. It is paired with a masked-reconstruction training objective and a latent-space variational resampling. The continual learning machinery is an ensemble of four such models, each trained on a different recent run and each applied to max-scaled and quantile-scaled inputs; an LS is flagged anomalous if any member's gap-score (the gap between the largest and second-largest p

What would settle it

Inject a known synthetic anomaly into the validation data used to recompute mu_err and sigma_err for a given run, then run the ensemble on that run's test data; if the injected anomaly is never flagged (recall near zero for that pattern), the pristine-data assumption would be falsified. A second test: replace the synthetic single-pixel multiplicative anomalies with a real detector fault signature (e.g., the 2018 dead region in a 40-degree azimuthal sector) and check whether the ensemble detects it as a single entity.

Watch

Extended reading notes

Core claim

DepthViT is a masked autoencoder with depth-wise convolutional patch embeddings and depth-wise attention, meaning each of the seven HCAL depth channels keeps its own kernels and attention is computed along the channel axis rather than the spatial patch axis. The authors report that on a single 2018 run a DepthViT sub-ensemble detects injected single-pixel anomalies with perfect precision and recall for most anomaly factors, but that the same static model degrades severely on later runs, reaching false-negative rates of 0.45–0.75 on 2022 data. Their continual learning method refreshes the per-pixel error statistics (mu_err, sigma_err) used in the Z-score baseline using each run's validation d

Load-bearing premise

The method relies on each new run providing a segment of 'pristine' (anomaly-free) data to recompute error statistics and train new ensemble members; if anomalies are present in that calibration data, they become part of the baseline and escape detection.

Editorial extensions

If this is right

  • Detector quality monitoring can stay accurate over multi-year campaigns by retraining only on recent runs and keeping a small ensemble, rather than retraining on all historical data.
  • The same continual learning recipe should apply to any monitoring setting where sensor readings drift with time, such as manufacturing lines or environmental sensors.
  • Because DepthViT uses about 300k parameters, ensembling many models adds little computational cost; the paper notes the OR-ensemble is trivially parallelizable.
  • The combination of refreshed statistics and ensembling specifically reduces false-negative rates (from 0.45–0.75 down to near zero for strong anomalies) while keeping false positives low (FPR 0.012).

Reading between the lines

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

  • The evaluation uses synthetic single-pixel multiplicative anomalies; a natural next test is whether the ensemble detects extended, correlated real faults such as the 2018 dead azimuthal sector, which may behave differently under max and quantile scaling.
  • The method requires a clean stretch of data for every run to refresh statistics; in practice, an online deployment might need a separate robust estimator of baseline statistics or a human-in-the-loop confirmation that the calibration window is anomaly-free.
  • Depth-wise attention trades away spatial mixing between patches within a layer; this design could generalize to other multi-channel data where channels are not aligned to the same physical point, such as spectral or multi-band sensor data—an extension the paper mentions but does not evaluate.
  • Using a logical OR ensemble makes the system sensitive to any single model's false positive, so the gain in recall comes with a structural risk; the reported FPR of 0.012 may depend on the gap-score threshold and the number of ensemble members, and both may need recalibration over time.
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 / 6 minor

Summary. The paper proposes DepthViT, a lightweight masked-autoencoder architecture that uses depth-wise convolutional embeddings and cross-depth attention, for anomaly detection on CMS HCAL DigiOccupancy maps. It combines two continual-learning mechanisms: (i) per-run recalculation of the per-pixel error statistics (mu_err, sigma_err) used in a Z-score detector, and (ii) an ensemble of DepthViT sub-models, each trained on a recent run, whose outputs are combined with a logical OR rule. Evaluated across 2018 and 2022 HCAL runs, the authors report that a static model degrades severely (Table 2), while the combined ensemble-plus-refreshed-statistics method maintains precision above 0.987 and low FNR across all tested anomaly factors (Table 5). A control study (Table 6 vs. Table 5) shows that ensembling improves recall for strong anomalies compared to using only the latest model.

Significance. If the results are robust, the paper offers a simple and computationally cheap continual-learning recipe for DQM that could transfer to other streaming anomaly-detection settings. The work has concrete strengths: hyperparameters are tuned on validation data, the comparison to a latest-model-only baseline is a useful control, the code is made available, and the paper is honest about its reliance on pristine data for recalibration. However, the evaluation is narrow — single-pixel synthetic anomalies, no contamination of the recalibration stream, no error bars — and the architecture's claimed superiority over existing DQM anomaly detectors is not benchmarked. These limitations materially affect the strength of the headline claims.

major comments (4)
  1. [Sec. 5.3, Eqs. (4.11)-(4.12)] The combined method's recalibration step depends on 'continued availability of pristine detector data' (stated in Sec. 5.3). In a live DQM stream, anomalies may appear in the very validation LS used to recompute mu_err and sigma_err; a persistent anomaly on a channel would shift the per-pixel baseline and suppress its Z-score, preventing the gap-score from firing. The paper provides no sensitivity analysis to validation-set contamination. A concrete test would be to contaminate a controlled fraction of validation LS with anomalies and report precision/recall vs. contamination level. Without such a test, Tables 5 and the abstract's claim of 'stable ratio of correct anomaly predictions' are conditional on an oracle-clean recalibration stream.
  2. [Sec. 4.2 (gap-score); Sec. 5.1 (synthetic anomalies)] All quantitative evaluation uses single-pixel, randomly located multiplicative anomalies. The gap-score in Eq. (4.12) uses only the two largest Z-scores, an explicit single-anomaly design. Real HCAL faults can be regional, as in the 40-degree power loss shown in Fig. 3(a). It is not shown that the method detects multi-pixel or regional anomalies; the extrapolation in Sec. 4.2 is untested. The claim that the method maintains a stable detection ratio should be restricted to single-pixel anomalies, or the evaluation should include regional/multi-pixel faults.
  3. [Tables 5-6, Sec. 5.4] Results are point estimates from a single data split and a single random anomaly-location realization. There are no error bars, confidence intervals, or repeated trials. The differences between the ensemble (Table 5) and latest-model-only (Table 6) — e.g., FNR 0.010 vs 0.022 at anomaly factor 1.5 — may or may not be statistically significant. Adding bootstrap or repeated-seed trials would allow the reader to judge the stability of the continual-learning benefit.
  4. [Introduction, Sec. 5] The paper claims 'competitive performance relative to current state-of-the-art approaches' for the DepthViT architecture, but no comparison to existing DQM anomaly detectors (e.g., the graph-network or transfer-learning baselines cited as refs. [6] and [8]) is reported. The continual-learning result stands on its own as a controlled comparison, but the architecture's claimed advantage over the state of the art is not demonstrated.
minor comments (6)
  1. [Abstract] The abstract in the submitted text states 'precision above 98.8%' while the full-text abstract says 'above 99%'. Please make consistent.
  2. [Eq. (4.11)] Notation is inconsistent (pred_err vs prederr) and the equation has ambiguous parentheses in the denominator. Please clean up the typesetting and define all symbols.
  3. [Sec. 3] The reported percentage changes (60% reduction, 188% increase) should be double-checked against the plotted mean DigiOccupancy values; the 188% increase description seems inconsistent with the downward trend within a run described in the same paragraph.
  4. [Figs. 8-9] The figure captions refer to 'pointers' without explaining what they indicate; improve caption clarity.
  5. [Sec. 4.1, Eq. (4.1)] The notation for the conventional convolution is hard to follow (h', w', h'', w'' are not clearly defined). Consider using standard index notation.
  6. [Reference [29]] Reference [29] is cited as the source for the 40-degree HCAL power loss, but the title 'Search for disappearing tracks...' appears unrelated. Please verify the citation or replace it with the appropriate technical report.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the DepthViT derivation and ensemble evaluation are self-contained; the 'pristine validation data' requirement is an explicit scope condition, not a circular step.

full rationale

The claimed derivation chain is not circular. DepthViT is trained on a train split of a single run; mu_err and sigma_err are computed by applying the model to a separate validation split (Sec 5.1); synthetic anomalies are injected only into the test split (Sec 4.2, Sec 5.1). The Z-score detector (Eq. 4.11) and gap-score rule (Eqs. 4.12, G0=0.3) are then evaluated on test data that was not used to compute the baseline statistics. The continual-learning results in Sec 5.3 are honest comparisons: updating mu/sigma from each run's validation set, training new ensemble members, and testing on that run's test set with injected anomalies. No fitted parameter is fed back as the predicted target. The explicit statement that the method 'relies solely on the continued availability of pristine detector data' (Sec 5.3) is a genuine limitation for live DQM streams, since contamination of the validation data would absorb anomalies into the baseline; however, this is an external assumption about deployment data quality, not a circular definition or a fitted quantity renamed as prediction. The threshold G0 is optimized once, which is a model-selection/generalization concern rather than circularity. There are no load-bearing self-citations or imported uniqueness theorems.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a synthetic-anomaly proxy, per-run pristine validation data, single-anomaly-per-LS detection, and a fixed 4-run ensemble window. These are stated in the paper but not independently verified; hyperparameters G0, T, patch size, kernels, mask ratio, layers, and latent dimension are all tuned on validation. No new physical entities are introduced.

free parameters (9)
  • G0 (gap-score threshold) = 0.3
    Gap-score anomaly threshold optimized on validation data according to anomaly detection performance (Sec 5.1).
  • Time window T = 5 lumisections
    Selected as a trade-off between transient and persistent anomaly recall (Sec 5.1).
  • Patch size = 12
    Hyperparameter chosen via optimization; varying patch size did not improve performance (Sec 5.1).
  • Number of kernels = 8
    Hyperparameter selected according to anomaly detection performance (Table 1 caption).
  • Mask ratio = 0.5
    MAE masking likelihood chosen via optimization (Table 1 caption).
  • Number of layers = 5
    Encoder/decoder depth chosen via optimization (Table 1 caption).
  • Latent dimension = 56
    Latent dimension chosen via optimization (Table 1 caption).
  • Ensemble size = 4 runs
    Fixed sliding window of four run sub-ensembles; no ablation or optimization is reported (Sec 5.3).
  • Quantile-scaling epsilon = unspecified small value
    Introduced in Eq. 4.13 to avoid divide-by-zero; its value is not reported.
assumptions (5)
  • domain assumption Synthetic single-pixel multiplicative anomalies (factors 0.0-2.0) faithfully mimic real HCAL dead, degraded, and hot channels.
    All quantitative claims in Tables 1-6 are generated from these injected anomalies (Sec 4.2, Sec 5.1); real anomaly data is not used.
  • domain assumption Pristine, anomaly-free validation data is available for every new run to update mu_err/sigma_err and to train new models.
    Section 5.3 explicitly states the mitigation technique relies on continued availability of pristine detector data; if the stream is already anomalous, recalibration absorbs the fault into the baseline.
  • domain assumption Only a single anomaly per lumisection needs to be detected.
    The gap-score uses only the two largest Z-scores and the authors state 'only single anomalies per LS' are considered (Sec 4.2, Eq. 4.12).
  • standard math Per-pixel prediction errors for pristine data are approximately independent and identically distributed so Eq. 4.11 follows a standard normal under the null.
    The Z-score formula (Eq. 4.11) relies on a CLT-style scaling of summed errors; no normality or autocorrelation test is provided.
  • domain assumption A sliding window of the four most recent runs is sufficient to span the relevant distribution-shift regime.
    The ensemble fixes the window at four runs without ablating window size or testing retention on older runs (Sec 5.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Continual Learning via Ensemble-Based Depth-Wise Masked Autoencoders for Data Quality Monitoring in High-Energy Physics." pith.science (2026). https://pith.science/paper/YQPYN3LO

@misc{pith2026260302369,
  author       = {Pith},
  title        = {Pith review of: Continual Learning via Ensemble-Based Depth-Wise Masked Autoencoders for Data Quality Monitoring in High-Energy Physics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YQPYN3LO}},
  note         = {Machine review of arXiv:2603.02369}
}
read the original abstract

Machine learning (ML) techniques have been demonstrated to improve the accuracy and efficiency of anomaly detection (AD) when compared to conventional methods. This has led to the adoption of ML for data quality monitoring (DQM) use cases in order to monitor the operation of certain systems to ensure that they are free of undesirable or potentially deleterious anomalies. For applications in the field of High-Energy physics (HEP), where detectors must operate in long-running, harsh environments, ML models used in DQM that have been trained on static datasets are bound to experience degraded performance due to distributional shifts that naturally occur in the incoming data streams, unless directly mitigated via the inclusion of continual ML techniques. This work introduces DepthViT, a lightweight masked autoencoder architecture that employs unique depth-wise embeddings and depth-wise attention, to perform computationally efficient AD tasks. A continual learning framework is developed in which DepthViT models trained on the most recent data streams are ensembled with older models to create a robust overall system which is more resilient to shifts in incoming data streams. When evaluated on occupancy maps from the Compact Muon Solenoid (CMS) hadron calorimeter across multiple data-taking campaigns, the proposed method maintains precision above 98.8\% and a stable ratio of correct anomaly predictions to number of anomalies both under small and large distributional shifts. Beyond HEP, the same ensembling-based continual adaptation strategy can be directly applied to industrial monitoring environments where data also naturally evolve over time. This work therefore presents a path toward adaptive anomaly detection systems capable of sustained operation in dynamic data environments.

Figures

Figures reproduced from arXiv: 2603.02369 by the authors.

Figure 1
Figure 1. (a) The CMS detector is comprised of four sub-detectors in addition to a 3.8T superconducting solenoid magnet. Each sub-detector specializes in the detection of different particle species originating from collision interactions. (b) The HCAL sub-detector consists of multiple regions, including the HE and HB regions. The depth and iη and iϕ (discretely index pseudorapidity and azimuthal angle) values of these regions… view at source ↗
Figure 2
Figure 2. Mean DigiOccupancy versus LS. The red vertical line delineates data originating from 2018 (left) and data originating 2022 (right). The data is temporally ordered both within and between runs, however it should not be assumed that run data is recorded in equally spaced intervals of time. approximately one week between runs). Each DigiOccupancy map is partitioned into 3D spatial coordinates denoted by iη ∈ [−32, 32],… view at source ↗
Figure 3
Figure 3. (a) DigiOccupancy map corresponding to Run325117, LS 0 (2018). Note the large absence of DigiOccupancy values in the upper left hand corner which resulted from a noted detector failure. (b) DigiOccupancy map corresponding to Run355456, LS 0 (2022). 4 METHODOLOGY 4.1 DepthViT Architecture Overview Here, we introduce a novel masked autoencoder architecture with depth-wise embeddings and cross-depth attention, referred… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: DepthViT variational autoencoder architecture which uses depthwise convolutional embeddings and depth-wise attention, latent space resampling of the encoder outputs. summing over the channel dimension and imposing a pixel-level cross channel symmetry, instead each chan…
Figure 6
Figure 6. Figure 6: (a) Normalized Z-score distribution of a well-performing DepthViT model when deployed on non-anomalous data. (b) Normalized Z-score distribution of a well-performing DepthViT model when deployed on anomalous data with anomaly factor of 0.0. (c) Normalized Z-score distr…
Figure 7
Figure 7. Figure 7: Flow chart detailing the ensembling scheme implemented within this work. Input data is scaled separately using max scaling and quantile scaling operations. The data is then fed in a parallel fashion into individual sub-ensembles which consist of independent DepthViT mo…
Figure 8
Figure 8. Figure 8: FPR (left) and FNR (right) of DepthViT model trained on Run323940 using the same hyperparameters detailed in Tab. 1 when evaluated on test data from subsequent runs. Pointers indicate which dataset the model was evaluated on, while the red dashed line delineates runs o…
Figure 9
Figure 9. Figure 9: FPR (left) and FNR (right) of the baseline DepthViT model trained on Run323940, using the same hyperparam￾eters detailed in Tab. 1 but with updated µerr and σerr values derived from deployment on the validation data of the dataset being evaluated. Pointers indicate whi…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 2 canonical work pages

  1. [6]

    Spatio-Temporal Anomaly Detection with Graph Networks for Data Quality Mon- itoring of the Hadron Calorimeter,

    M. W. Asres et al., “Spatio-Temporal Anomaly Detection with Graph Networks for Data Quality Mon- itoring of the Hadron Calorimeter,”Sensors, vol. 23, no. 24, 2023,ISSN: 1424-8220.DOI:10.3390/ s23249679

    2023

  2. [8]

    Data Quality Monitoring for the Hadron Calorimeters Using Transfer Learning for Anomaly Detection,

    M. W. Asres et al., “Data Quality Monitoring for the Hadron Calorimeters Using Transfer Learning for Anomaly Detection,”Sensors, vol. 25, no. 11, 2025,ISSN: 1424-8220.DOI:10.3390/s25113475

    work page doi:10.3390/s25113475 2025

  3. [1]

    Scientific discovery in the age of artificial intelligence,

    H. Wang et al., “Scientific discovery in the age of artificial intelligence,”Nature, vol. 620, no. 7972, pp. 47–60, 2023,ISSN: 1476-4687.DOI:10.1038/s41586-023-06221-2 16

    doi:10.1038/s41586-023-06221-2 2023

  4. [2]

    A Comprehensive Investigation of Anomaly Detection Methods in Deep Learning and Machine Learning: 2019–2023,

    S. Kumari, C. Prabha, A. Karim, M. M. Hassan, and S. Azam, “A Comprehensive Investigation of Anomaly Detection Methods in Deep Learning and Machine Learning: 2019–2023,”IET Information Security, vol. 2024, no. 1, p. 8 821 891, 2024.DOI:https : / / doi . org / 10 . 1049 / 2024 / 8821891

    2019

  5. [3]

    The Data Quality Monitoring Software for the CMS experiment at the LHC: past, present and future,

    Azzolini, Virginia et al., “The Data Quality Monitoring Software for the CMS experiment at the LHC: past, present and future,”EPJ Web Conf., vol. 214, p. 02 003, 2019.DOI:10 . 1051 / epjconf / 201921402003

    2019

  6. [4]

    Machine Learning Anomaly Detection Applications to Compact Muon Solenoid Data Qual- ity Monitoring,

    A. A. Pol, “Machine Learning Anomaly Detection Applications to Compact Muon Solenoid Data Qual- ity Monitoring,” Theses, Universit´e Paris-Saclay, Jun. 2020. [Online]. Available:https://theses. hal.science/tel-02924477

    2020

  7. [5]

    Machine learning for anomaly detection in particle physics,

    V . Belis, P. Odagiu, and T. K. Aarrestad, “Machine learning for anomaly detection in particle physics,” Reviews in Physics, vol. 12, p. 100 091, 2024,ISSN: 2405-4283.DOI:https : / / doi . org / 10 . 1016/j.revip.2024.100091

    arXiv 2024

  8. [7]

    Long Horizon Anomaly Prediction in Multivariate Time Series with Causal Autoen- coders,

    M. W. Asres et al., “Long Horizon Anomaly Prediction in Multivariate Time Series with Causal Autoen- coders,”PHM Society European Conference, vol. 7, no. 1, pp. 21–31, 2022.DOI:10.36001/phme. 2022.v7i1.3367

    work page doi:10.36001/phme 2022

Show all 33 references
  1. [9]

    The upgrade of the CMS hadron calorimeter with silicon photomultipliers,

    N. Strobbe, “The upgrade of the CMS hadron calorimeter with silicon photomultipliers,”Journal of Instrumentation, vol. 12, no. 01, p. C01080, Jan. 2017.DOI:10 . 1088 / 1748 - 0221 / 12 / 01 / C01080

    2017

  2. [10]

    LHC Machine,

    L. Evans and P. Bryant, “LHC Machine,”Journal of Instrumentation, vol. 3, no. 08, S08001, Aug. 2008. DOI:10.1088/1748-0221/3/08/S08001

    2008 doi

  3. [11]

    The CMS experiment at the CERN LHC,

    CMS Collaboration, “The CMS experiment at the CERN LHC,”Journal of Instrumentation, vol. 3, no. 08, S08004, Aug. 2008.DOI:10.1088/1748-0221/3/08/S08004

    2008 doi

  4. [12]

    CMS Technical Design Report for the Phase 1 Upgrade of the Hadron Calorimeter,

    J. Mans et al., “CMS Technical Design Report for the Phase 1 Upgrade of the Hadron Calorimeter,” CERN, Tech. Rep., 2012. [Online]. Available:https://cds.cern.ch/record/1481837

    2012

  5. [13]

    The CMS trigger system,

    CMS Collaboration, “The CMS trigger system,”Journal of Instrumentation, vol. 12, no. 01, P01020, Jan. 2017.DOI:10.1088/1748-0221/12/01/P01020

    2017 doi

  6. [14]

    Characteristics of the ATLAS and CMS detectors,

    A. Seiden, “Characteristics of the ATLAS and CMS detectors,”Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, vol. 370, pp. 892–906, 2012.DOI:10. 1098/rsta.2011.0461

    2012

  7. [15]

    CMS data quality monitoring: Systems and experi- ences,

    L. Tuura, A. Meyer, I. Segoni, and G. D. Ricca, “CMS data quality monitoring: Systems and experi- ences,”Journal of Physics: Conference Series, vol. 219, no. 7, p. 072 020, Apr. 2010.DOI:10.1088/ 1742-6596/219/7/072020

    2010

  8. [16]

    Overcoming catastrophic forgetting in neural networks,

    J. Kirkpatrick et al., “Overcoming catastrophic forgetting in neural networks,”Proceedings of the Na- tional Academy of Sciences, vol. 114, no. 13, pp. 3521–3526, 2017.DOI:10.1073/pnas.1611835114 17

    2017 doi

  9. [17]

    Continual learning through synaptic intelligence,

    F. Zenke, B. Poole, and S. Ganguli, “Continual learning through synaptic intelligence,” inProceed- ings of the 34th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 70, PMLR, 2017, pp. 3987–3995

    2017

  10. [18]

    Gradient episodic memory for continual learning,

    D. Lopez-Paz and M. Ranzato, “Gradient episodic memory for continual learning,” inAdvances in Neural Information Processing Systems, vol. 30, 2017, pp. 6467–6476

    2017

  11. [19]

    Icarl: Incremental classifier and represen- tation learning,

    S.-A. Rebuffi, A. Kolesnikov, G. Sperl, and C. H. Lampert, “Icarl: Incremental classifier and represen- tation learning,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 5533–5542.DOI:10.1109/CVPR.2017.587

    2017 doi

  12. [20]

    Progressive neural networks,

    A. A. Rusu et al., “Progressive neural networks,”arXiv preprint arXiv:1606.04671, 2016

    2016 arXiv

  13. [21]

    Packnet: Adding multiple tasks to a single network by iterative pruning,

    A. Mallya and S. Lazebnik, “Packnet: Adding multiple tasks to a single network by iterative pruning,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 7765– 7773.DOI:10.1109/CVPR.2018.00810

    2018

  14. [22]

    I. E. Marouf, S. Roy, E. Tartaglione, and S. Lathuili `ere,Weighted Ensemble Models Are Strong Contin- ual Learners, 2024. arXiv:2312.08977 [cs.LG]. [Online]. Available:https://arxiv.org/ abs/2312.08977

    2024 arXiv

  15. [23]

    T. Doan, S. I. Mirzadeh, and M. Farajtabar,Continual Learning Beyond a Single Model, 2023. arXiv: 2202.09826 [cs.LG]. [Online]. Available:https://arxiv.org/abs/2202.09826

    2023 arXiv

  16. [24]

    Zhou, H.-L

    D.-W. Zhou, H.-L. Sun, H.-J. Ye, and D.-C. Zhan,Expandable Subspace Ensemble for Pre-Trained Model-Based Class-Incremental Learning, 2024. arXiv:2403.12030 [cs.CV]. [Online]. Available: https://arxiv.org/abs/2403.12030

    2024 arXiv

  17. [25]

    W ´ojcik, W

    M. W ´ojcik, W. Ko´sciukiewicz, T. Kajdanowicz, and A. Gonczarek,Neural Architecture for Online En- semble Continual Learning, 2023. arXiv:2211 . 14963 [cs.LG]. [Online]. Available:https : //arxiv.org/abs/2211.14963

    2023 arXiv

  18. [26]

    Rype ´s´c, S

    G. Rype ´s´c, S. Cygert, V . Khan, T. Trzci´nski, B. Zieli ´nski, and B. Twardowski,Divide and not forget: Ensemble of selectively trained experts in Continual Learning, 2024. arXiv:2401.10191 [cs.LG]. [Online]. Available:https://arxiv.org/abs/2401.10191

    2024 arXiv

  19. [27]

    Ye and A

    F. Ye and A. G. Bors,Learning an evolved mixture model for task-free continual learning, 2022. arXiv: 2207.05080 [cs.LG]. [Online]. Available:https://arxiv.org/abs/2207.05080

    2022 arXiv

  20. [28]

    O. S. Br ¨uning et al.,LHC Design Report(CERN Yellow Reports: Monographs). Geneva: CERN, 2004. DOI:10.5170/CERN-2004-003-V-1

    2004 doi

  21. [29]

    Search for disappearing tracks in proton-proton collisions at √s= 13 TeV,

    The CMS Collaboration, “Search for disappearing tracks in proton-proton collisions at √s= 13 TeV,” Physics Letters B, vol. 806, p. 135 502, 2020,ISSN: 0370-2693.DOI:https : / / doi . org / 10 . 1016/j.physletb.2020.135502

    2020

  22. [30]

    Dosovitskiy et al.,An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,

    A. Dosovitskiy et al.,An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,

  23. [31]

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick,Masked Autoencoders Are Scalable Vision Learners, 2021. arXiv:2111 . 06377 [cs.CV]. [Online]. Available:https : / / arxiv . org / abs/2111.06377 18

    2021 arXiv

  24. [32]

    Ristea, F.-A

    N.-C. Ristea, F.-A. Croitoru, R. T. Ionescu, M. Popescu, F. S. Khan, and M. Shah,Self-distilled masked auto-encoders are efficient video anomaly detectors, 2024. arXiv:2306.12041 [cs.CV]. [Online]. Available:https://arxiv.org/abs/2306.12041 19 The CMS HCAL Collaboration Yereva...

    2024 arXiv

  25. [2021]

    [Online]

    arXiv:2010.11929 [cs.CV]. [Online]. Available:https://arxiv.org/abs/2010. 11929

    2010 arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.