Pith. sign in

REVIEW 3 major objections 6 minor 36 references

WhACC: Whisker Automatic Contact Classifier with Expert Human-Level Performance

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read WhACC, a two-stage hybrid of a convolutional neural network and gradient-boosted trees, labels whisker touches in high-speed video with expert human-level accuracy, and its retraining workflow cuts the human curation of a…

desk verdict Well-engineered, honestly reported tool; the human-level claim holds for the main test set, but the retraining validation would be stronger with a human-error baseline. read the letter →

arxiv 2501.06219 v1 pith:ZZMJJ5AG submitted 2025-01-06 cs.CV cs.LG

Phillip Maire , Samson G. King , Jonathan Andrew Cheung , Stefanie Walker , Samuel Andrew Hires This is my paper · ORCID
classification cs.CVcs.LG
keywords WhACCwhiskertouchclassificationhigh-speedvideoResNet50V2LightGBMcounterrortemporalfeatureengineeringvibrissalsystem
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 claims that WhACC, a two-stage model built from a ResNet50V2 feature extractor and a LightGBM classifier, can label whisker-pole touches in 1,000 fps video as accurately as expert human curators, in a fully automated pass. On test data spanning two labs and eight years, WhACC agreed with expert curators on about 99.5% of frames and made fewer touch-count errors on average than any individual curator, though it committed more edge errors and therefore tended to shorten touches slightly. The authors argue this matters because manual curation of touch events currently costs over three hours per million video frames, and their retraining workflow, validated on 16 held-out sessions of about four million frames, cuts the human time to roughly six hours for a 100-million-frame dataset. The goal is to replace the second-stage manual curation that current tracking tools still require, giving systems neuroscience millisecond-precision touch labels from video alone.

What carries the argument

The load-bearing mechanism is the two-stage hybrid WhACC pipeline. Stage one is a fully unfrozen ResNet50V2, pretrained on ImageNet and trained on 61x61 pixel crops centered on the object, with three consecutive video frames stacked into the cyan, magenta, and yellow color channels so each prediction has access to times t, t-1, and t-2; training uses random rotation, shift, zoom, brightness, and Gaussian noise augmentation. Stage two feeds the 2,048 penultimate-layer features through engineered temporal transformations, including forward and backward shifts up to five frames, rolling means and standard deviations over windows from 3 to 61 frames, and discrete differences over offsets from -50 to +50, yielding 84,009 candidate features that an ensemble of LightGBM models recursively prunes to 2,105 high-value features. A final LightGBM classifier, tuned by Optuna and followed by five-frame median smoothing, makes the touch or non-touch decision. These two temporal mechanisms, lag-frame color stacking and engineered rolling and difference features, are what let a per-frame CNN classifier behave like a curator who scrolls back and forth to find touch onset and offset.

What would settle it

Mount a force or strain sensor on the pole, synchronized to the video, so physical contact times are known independently; then compare WhACC's and a human curator's touch-count and edge errors against that sensor. If WhACC's touch-count errors relative to the sensor are significantly worse than the human curator's, the 'expert human-level' claim would be refuted.

Watch

Extended reading notes

Core claim

The central discovery is that expert-level touch classification does not require a larger or more complex network: a comparatively lightweight CNN trained on small object-centered crops, followed by engineered temporal features and gradient-boosted trees, matches human curation while being fully automatic. The paper demonstrates this by comparing WhACC against three expert curators on a shared test set, using a majority-of-three consensus as ground truth. Measured by touch-count errors (splits, ghosts, misses, joins), the errors that most degrade touch-count and peri-stimulus time histogram analyses, WhACC outperformed the average human curator; measured by edge errors it made more deductions, trimming touches at their start. Pairwise frame agreement was about 99.5%, equal to agreement between human curators. A separate 16-session holdout showed that retraining LightGBM on 100 to 1,000 frames per session rescues poorly performing sessions, and touch-evoked PSTHs from WhACC labels matched a human curator's for touch-responsive neurons.

Load-bearing premise

The load-bearing premise is that the majority vote of the paper's three expert human curators is an adequate ground truth for touch; if those humans are systematically biased or noisy, especially around touch onset and offset, then 'expert human-level' means agreement with those annotators rather than true detection.

Editorial extensions

If this is right

  • For labs using similar single-whisker, pole-contact, 1,000 fps head-fixed setups, WhACC can serve as a fully automated first-pass curator, with touch-count errors at or below the average human curator.
  • Because WhACC's PSTHs matched human curators for spike counts on seven touch-responsive neurons, its labels can feed standard electrophysiology analyses without manual re-curation, though peak responses ran one to two milliseconds early.
  • The retraining workflow, validated on 16 held-out sessions of about four million frames, fixed most of the five initially poor sessions with 100 frames per session and produced additional gains with 1,000 frames per session.
  • Median smoothing of predictions is a simple uniform booster: the paper reports it alone cut touch-count error by 80% in the original-feature model, and combining smoothing with feature engineering reduced touch-count error to 0.202.
  • For datasets outside the training conditions, such as different frame rates, multi-whisker video, or non-pole objects, performance is not established without retraining.

Reading between the lines

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

  • A natural next test is to instrument the pole with a force or strain sensor to create an independent physical ground truth; if human labels are themselves noisy at onset and offset, WhACC's consistency could actually exceed human performance for population-level measures.
  • The recipe of CNN features plus engineered temporal features plus gradient boosting plus median smoothing is not whisker-specific, and could be reused for other high-speed video tasks where small object-contact events must be labeled with temporal consistency.
  • The 100-frame retraining result suggests a cheaper active-learning loop: sample frames near predicted touch onsets and offsets, curate only those, and retrain, which could reduce the six-hour estimate further.
  • Because lag images hurt MobileNetV3 models, putting temporal information in color channels interacts with architecture choice; engineered features on CNN embeddings may be a more transferable way to add temporal context than recurrent layers.
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 / 6 minor

Summary. The manuscript describes WhACC, a two-stage pipeline (ResNet50V2 feature extraction followed by LightGBM classification) for detecting whisker-pole contact in head-fixed mouse high-speed video. Training labels are majority decisions of three human curators; inputs are object-centered 61x61 crops with temporal lag frames and augmentation. The authors report that WhACC achieves human-level or better touch-count error on a large test set, that a feature-selection procedure reduces 84,009 features to 2,105, and that a small-sample retraining protocol improves performance on 16 held-out sessions (~4 million frames). The claimed practical benefit is a reduction from about 333 to about 6 curation hours per 100 million frames.

Significance. WhACC addresses a real bottleneck in systems neuroscience, and the paper has several strengths: the model comparison is extensive, the code and model are promised in a public repository, limitation sections are candid, and the retraining idea is pragmatic. If the human-level claim is established on a properly independent benchmark, this would be a useful methods contribution. The main risks are that part of the reported feature-selection gain is evaluated on data used to choose features, and the retraining validation lacks an independent reference label set; both need to be addressed before the headline performance claims can be taken at face value.

major comments (3)
  1. [§1.4 and §3.4] The feature-selection procedure is not leakage-free. The manuscript states in §1.4 that "some test data was used for feature selection" and in §3.4 that the split for feature selection was made "based on frame index (as opposed to segment or video)". Because the engineered features include rolling windows up to 61 frames, shifts, and discrete differences up to ±50 frames, a frame-index split allows temporal context from training frames to enter validation features. This is load-bearing: Table 1 reports the selected-feature model as the best TC-error model (0.202 after smoothing vs. 0.214 for the full-feature model), and the final WhACC model uses the selected features. The reported advantage of feature selection, and possibly the Figure 5 human-level evaluation, is therefore not a clean generalization estimate. The authors should repeat feature selection on a split that respects video identity and does not touch the final test set, and report the independently selected model's performance on a fully untouched holdout.
  2. [§3.6 and Figure 6] The retraining validation uses a single expert curator as the reference, and the paper explicitly notes that "the human error rate is not known" for this dataset. Yet Figure 6 and the PSTH comparisons treat this single curator as ground truth for TC-error, spike counts, peak response time, and signal-window width. Since §1.2 shows that expert curators disagree on roughly 1 in 200 frames and that human touch-count errors are a measurable fraction of touches, the post-retraining error and the reported 1-2 ms earlier peak response cannot be attributed cleanly to the model. The pre/post retraining comparison does demonstrate improvement, but the claim of "expert-level performance on new data" requires a multi-curator consensus on at least a subset of the 16 sessions, or a re-analysis that explicitly accounts for the single curator's error rate.
  3. [§1.2 and §1.4] The comparison between WhACC and human curators is asymmetric in a way that can favor WhACC. The model is trained on majority labels from the same three curators whose pairwise consensus is used as the reference when estimating human error rates. The human error rate is estimated by holding one curator out from the two-curator consensus, but WhACC has been optimized to reproduce the majority behavior of all three curators. Even on unseen videos, consistent annotator biases can therefore be baked into WhACC's training target, making it easier for WhACC to match a two-curator consensus than a genuinely independent human would be. The comparison would be fairer if WhACC's training labels excluded the pair of curators used as the evaluation consensus, or if an independently generated annotation protocol were used to define the test-set reference.
minor comments (6)
  1. [Figure 6 caption] The caption states that gray outlines indicate sessions from the top row in A and black outlines indicate those from the top row in A; one of these should refer to the bottom row.
  2. [§2.6 and Table 1] The text refers to the "original-feature model with 2,408 features," but the correct number is 2,048; Table 1 also lists the full feature count as 84,006 while the text consistently says 84,009.
  3. [Abstract and §1.2] The abstract reports "pairwise touch classification agreement on 99.5% of video frames" while the body reports that all three curators agree on 99.46%; please define both quantities and reconcile the numbers.
  4. [Figure 5] The claim that WhACC made "fewer TC-errors on average" than human curators is not accompanied by an inferential test or the per-comparison values; with only three curators, a paired effect-size statement or exact comparisons would be useful.
  5. [§2.3] The statement that low-frame-rate tests "show promise" is anecdotal unless data are provided; please either add quantitative results or clearly label this as untested.
  6. [Throughout] The manuscript has many typographic artifacts (for example, "T able", "T o", "T rue") that should be corrected in copy editing.

Circularity Check

1 steps flagged · score 4.0 of 10

Admitted feature-selection leakage makes part of the benchmark comparison fitted to the test data; the human-level benchmark is otherwise a valid supervised target, and the 16-session holdout provides partial independent grounding.

  1. fitted input called prediction [Section 1.4, note after Figure 3C; Methods 3.5 (Preprocessing and training)]
    "Note that some test data was used for feature selection, thus the selected-feature model has some data leakage by way of which features were included. There is no leakage in the full-feature model."

    The 2,105-feature set is a fitted component of the pipeline, selected with the aid of test-set performance, and the same features are later used to train the final WhACC model. The final test set is a new random split from pooled sessions, and the paper does not show that frames used in feature selection are excluded from it. Reporting TC-errors for the selected-feature model as a model comparison is therefore a fitted-input-called-prediction: the feature selector saw the benchmark, and the benchmark score is presented as predictive performance.

full rationale

Most of WhACC's derivation is not circular: training a classifier on human majority labels and testing on held-out frames is exactly the stated task of automating human curation, and the 16-session holdout was not used in initial training, early stopping, or feature selection. The one clear fitted-input-called-prediction is the feature-selection stage: the paper admits test data were used to choose the 2,105 features, and the same features are used by the final model, with the final test split drawn from pooled sessions rather than explicitly separated from the feature-selection data. This makes the selected-feature model's Table 1 numbers partly fitted to the benchmark. In addition, the human-level comparison is constructed from the same three-curator majority used for training, so WhACC's agreement with pairwise consensus is a model-to-training-target measurement rather than a fully independent external benchmark; this is a benchmark-validity caveat but does not reduce the model's outputs to its inputs by construction because the test frames are held out. The retraining validation uses a single curator as reference and cannot separate model error from curator error, but that is a stated limitation rather than a circular step. Overall, the central claim retains substantial independent content, especially the retraining workflow and the time-saving result, but the reported model-comparison numbers are partially contaminated by the disclosed feature-selection leakage.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

No new physical entities are introduced. The load-bearing choices are model hyperparameters, feature-selection thresholds, data-selection decisions, and the use of human majority labels as ground truth.

free parameters (6)
  • ResNet50V2 CNN weights = fine-tuned on ~676,000 training frames
    Core feature extractor; trained on augmented images and lag frames, and selected based on test TC-error.
  • LightGBM classifier hyperparameters = optimized over 100 Optuna trials
    Regularization, number of leaves, feature fraction, bagging fraction, and minimum leaf data were tuned on validation TC-error and AUC.
  • Feature selection thresholds = 2,105 features retained from 84,009
    Thresholds for gain and split importance were chosen by trial and error during recursive elimination; the paper admits some test data was used in this process.
  • Median smoothing window = 5 frames
    Applied to all predictions; uniformly reduced TC-error across models, but is a chosen post-processing parameter.
  • Retraining sample weight = 2
    New sample data is weighted by a factor of two to bias the LightGBM model toward the new dataset, as described in Section 3.6.
  • Augmentation parameters = rotation, up to 10% shift, up to 25% zoom, 20-120% brightness, Gaussian noise scale 3
    Chosen by trial and error to keep frames interpretable to human curators; shapes the training distribution.
assumptions (6)
  • domain assumption Human majority labels are a valid proxy for ground truth touch status.
    Section 1.2 states no independent ground truth exists; training and evaluation use the majority of three expert curators.
  • domain assumption The object-centered 61x61 pixel crop contains sufficient information for touch classification.
    Section 1.1 crops video around the pole to exclude fur and other whiskers, assuming the relevant touch information is local to the object.
  • domain assumption Temporal context encoded by lag images and engineered rolling, shifting, and difference features captures the information used by human curators.
    Section 1.1 and Section 1.4 motivate lag images and feature engineering based on curator strategies such as scrolling between frames and inferring touch onset from displacement changes.
  • domain assumption ImageNet-pretrained CNN features transfer to grayscale whisker videos.
    Section 3.2 initializes all CNNs with ImageNet weights and fine-tunes them fully, relying on transfer learning from natural images to this domain.
  • domain assumption The test video sets from other labs and the 16-session holdout are representative of deployment conditions.
    Generalization claims rest on the representativeness of these datasets, as described in Section 1.3 and Section 3.6.
  • domain assumption Touch count error (TC-error) is the correct metric to prioritize for downstream neural analyses.
    Section 1.2 states that touch count errors most strongly degrade measures like total touch count and spike counts, and model selection is based solely on TC-error.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WhACC: Whisker Automatic Contact Classifier with Expert Human-Level Performance." pith.science (2026). https://pith.science/paper/ZZMJJ5AG

@misc{pith2026250106219,
  author       = {Pith},
  title        = {Pith review of: WhACC: Whisker Automatic Contact Classifier with Expert Human-Level Performance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZZMJJ5AG}},
  note         = {Machine review of arXiv:2501.06219}
}
read the original abstract

The rodent vibrissal system is pivotal in advancing neuroscience research, particularly for studies of cortical plasticity, learning, decision-making, sensory encoding, and sensorimotor integration. Despite the advantages, curating touch events is labor intensive and often requires >3 hours per million video frames, even after leveraging automated tools like the Janelia Whisker Tracker. We address this limitation by introducing Whisker Automatic Contact Classifier (WhACC), a python package designed to identify touch periods from high-speed videos of head-fixed behaving rodents with human-level performance. WhACC leverages ResNet50V2 for feature extraction, combined with LightGBM for Classification. Performance is assessed against three expert human curators on over one million frames. Pairwise touch classification agreement on 99.5% of video frames, equal to between-human agreement. Finally, we offer a custom retraining interface to allow model customization on a small subset of data, which was validated on four million frames across 16 single-unit electrophysiology recordings. Including this retraining step, we reduce human hours required to curate a 100 million frame dataset from ~333 hours to ~6 hours.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 34 canonical work pages

  1. [1]

    Whisking

    Sofroniew NJ, Svoboda K. Whisking. Current biology: CB 2015 Feb;25(4):R137–140

    work page 2015

  2. [2]

    Whisker touch sensing guides locomotion in small, quadrupedal mammals

    Grant RA, Breakell V, Prescott TJ. Whisker touch sensing guides locomotion in small, quadrupedal mammals. Proceedings Biological Sciences 2018 Jun;285(1880):20180592

    work page 2018

  3. [3]

    Vocalization-whisking coordination and multisensory integration of social signals in rat auditory cortex

    Rao RP , Mielke F, Bobrov E, Brecht M. Vocalization-whisking coordination and multisensory integration of social signals in rat auditory cortex. eLife 2014 Dec;3:e03185

    work page 2014

  4. [4]

    The sensorimotor basis of whisker-guided anteroposterior object localization in head-fixed mice

    Cheung J, Maire P , Kim J, Sy J, Hires SA. The sensorimotor basis of whisker-guided anteroposterior object localization in head-fixed mice. Current biology: CB 2019 Sep;29(18):3029–3040.e4

    work page 2019

  5. [5]

    Independent representations of self-motion and object location in barrel cortex output

    Cheung JA, Maire P , Kim J, Lee K, Flynn G, Hires SA. Independent representations of self-motion and object location in barrel cortex output. PLoS biology 2020 Nov;18(11):e3000882. Phillip Maire et al. 19

    work page 2020

  6. [6]

    Behavioral and neural bases of tactile shape discrimination learning in head-fixed mice

    Kim J, Erskine A, Cheung JA, Hires SA. Behavioral and neural bases of tactile shape discrimination learning in head-fixed mice. Neuron 2020 Dec;108(5):953–967.e8

    work page 2020

  7. [7]

    Sparse temporal coding of elementary tactile features during active whisker sensation

    Jadhav SP , Wolfe J, Feldman DE. Sparse temporal coding of elementary tactile features during active whisker sensation. Nature Neuroscience 2009 Jun;12(6):792–800. https://www.nature.com/articles/nn.2328

    work page 2009

  8. [8]

    Low-noise encoding of active touch by layer 4 in the somatosensory cortex

    Andrew Hires S, Gutnisky DA, Yu J, O’Connor DH, Svoboda K. Low-noise encoding of active touch by layer 4 in the somatosensory cortex. eLife 2015 Aug;4:e06619. https://doi.org/10.7554/eLife.06619

    work page doi:10.7554/elife.06619 2015

Show all 36 references
  1. [9]

    Organization of sensory feature selectivity in the whisker system

    Bale MR, Maravall M. Organization of sensory feature selectivity in the whisker system. Neuroscience 2018 Jan;368:70– 80

    2018

  2. [10]

    A General Classifier of Whisker Data Using Stationary Naive Bayes: Application to BIOTACT Robots

    Lepora NF, Fox CW, Evans M, Mitchinson B, Motiwala A, Sullivan JC, et al. A General Classifier of Whisker Data Using Stationary Naive Bayes: Application to BIOTACT Robots. In: Groß R, Alboul L, Melhuish C, Witkowski M, Prescott TJ, Penders J, editors. T owards Autonomous Roboti...

    2011

  3. [11]

    Automated tracking of whiskers in videos of head fixed rodents

    Clack NG, O’Connor DH, Huber D, Petreanu L, Hires A, Peron S, et al. Automated tracking of whiskers in videos of head fixed rodents. PLOS Computational Biology 2012 Jul;8(7):e1002591. https://journals.plos.org/ploscompbiol/ article?id=10.1371/journal.pcbi.1002591

    2012 doi

  4. [12]

    T racking whisker and head movements in unrestrained behaving rodents

    Knutsen PM, Derdikman D, Ahissar E. T racking whisker and head movements in unrestrained behaving rodents. Journal of Neurophysiology 2005 Apr;93(4):2294–2301

    2005

  5. [13]

    Unsupervised whisker tracking in unrestrained behaving animals

    Voigts J, Sakmann B, Celikel T. Unsupervised whisker tracking in unrestrained behaving animals. Journal of Neurophys- iology 2008 Jul;100(1):504–515

    2008

  6. [14]

    Unsupervised quantification of whisking and head movement in freely moving rodents

    Perkon I, Kosir A, Itskov PM, T asic J, Diamond ME. Unsupervised quantification of whisking and head movement in freely moving rodents. Journal of Neurophysiology 2011 Apr;105(4):1950–1962

    2011

  7. [15]

    The morphology of the rat vibrissal array: a model for quantifying spatiotemporal patterns of whisker-object contact

    T owal RB, Quist BW, Gopal V, Solomon JH, Hartmann MJZ. The morphology of the rat vibrissal array: a model for quantifying spatiotemporal patterns of whisker-object contact. PLoS computational biology 2011 Apr;7(4):e1001120

    2011

  8. [16]

    Whiskeras: a new algorithm for accurate whisker tracking

    Betting JHLF, Romano V, Al-Ars Z, Bosman LWJ, Strydis C, De Zeeuw CI. Whiskeras: a new algorithm for accurate whisker tracking. Frontiers in Cellular Neuroscience 2020;14:588445

    2020

  9. [17]

    Identity Mappings in Deep Residual Networks

    He K, Zhang X, Ren S, Sun J. Identity Mappings in Deep Residual Networks. In: Leibe B, Matas J, Sebe N, Welling M, editors. Computer Vision – ECCV 2016 Cham: Springer International Publishing; 2016. p. 630–645

    2016

  10. [18]

    Rethinking the Inception Architecture for Computer Vision

    Szegedy C, Vanhoucke V, Ioffe S, Shlens J, Wojna Z. Rethinking the Inception Architecture for Computer Vision. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR); 2016. p. 2818–2826

    2016

  11. [19]

    Searching for MobileNetV3

    Howard A, Sandler M, Chen B, Wang W, Chen LC, T an M, et al. Searching for MobileNetV3. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV); 2019. p. 1314–1324

    2019

  12. [20]

    Lightgbm: a highly efficient gradient boosting decision tree

    Ke G, Meng Q, Finley T, Wang T, Chen W, Ma W, et al. Lightgbm: a highly efficient gradient boosting decision tree. Advances in Neural Information Processing Systems 2017;30

    2017

  13. [21]

    Optuna: a next-generation hyperparameter optimization framework

    Akiba T, Sano S, Y anase T, Ohta T, Koyama M. Optuna: a next-generation hyperparameter optimization framework. In: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining KDD ’19, Anchorage, AK, USA: Association for Computing Machinery;...

    2019 doi

  14. [22]

    Xception: Deep Learning with Depthwise Separable Convolutions

    Chollet F . Xception: Deep Learning with Depthwise Separable Convolutions. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Honolulu, HI: IEEE; 2017. p. 1800–1807. 20 Phillip Maire et al

    2017

  15. [23]

    Depthwise convolution is all you need for learning multiple visual domains

    Guo Y, Li Y, Feris R, Wang L, Rosing T. Depthwise convolution is all you need for learning multiple visual domains. arXiv:190200927 [cs] 2019 Feb;http://arxiv.org/abs/1902.00927, arXiv: 1902.00927

    2019 arXiv

  16. [24]

    Rethinking Spatiotemporal Feature Learning: Speed-Accuracy T rade-offs in Video Classification

    Xie S, Sun C, Huang J, T u Z, Murphy K. Rethinking Spatiotemporal Feature Learning: Speed-Accuracy T rade-offs in Video Classification. In: Proceedings of the European Conference on Computer Vision (ECCV); 2018

    2018

  17. [25]

    Extreme Gradient Boost with CNN: A Deep Learning-Based Approach for Predicting Protein Subcellular Localization

    Ismail M, Islam Mondal MN. Extreme Gradient Boost with CNN: A Deep Learning-Based Approach for Predicting Protein Subcellular Localization. In: Arefin MS, Kaiser MS, Bandyopadhyay A, Ahad MAR, Ray K, editors. Proceedings of the International Conference on Big Data, Io T, and Ma...

    2022

  18. [26]

    Integration of convolutional neural network and extreme gradient boosting for breast cancer detection

    Sugiharti E, Arifudin R, Wiyanti DT, Susilo AB. Integration of convolutional neural network and extreme gradient boosting for breast cancer detection. Bulletin of Electrical Engineering and Informatics 2022 Apr;11(2):803–813. https://beei. org/index.php/EEI/article/view/3562

    2022

  19. [27]

    Gradient Boost with Convolution Neural Network for Stock Forecast

    Liu J, Lin C, Chao F . Gradient Boost with Convolution Neural Network for Stock Forecast. CoRR 2019;abs/1909.09563. http://arxiv.org/abs/1909.09563

    2019 arXiv

  20. [28]

    A model combining convolutional neural network and lightgbm algorithm for ultra-short-term wind power forecasting

    Ju Y, Sun G, Chen Q, Zhang M, Zhu H, Rehman MU. A model combining convolutional neural network and lightgbm algorithm for ultra-short-term wind power forecasting. IEEE Access 2019;7:28309–28318. https://ieeexplore.ieee. org/document/8653826/

    2019

  21. [29]

    Acoustic Scene Classification by Ensembling Gradient Boosting Machine and Convolutional Neural Networks; 2017

    Fonseca E, Gong R, Bogdanov D, Slizovskaia O, Gómez Gutiérrez E, Serra X. Acoustic Scene Classification by Ensembling Gradient Boosting Machine and Convolutional Neural Networks; 2017

    2017

  22. [30]

    A deep learning architecture for temporal sleep stage classifi- cation using multivariate and multimodal time series

    Chambon S, Galtier MN, Arnal PJ, Wainrib G, Gramfort A. A deep learning architecture for temporal sleep stage classifi- cation using multivariate and multimodal time series. IEEE transactions on neural systems and rehabilitation engineering: a publication of the IEEE Engineerin...

    2018

  23. [31]

    ConvXGB: A new deep learning model for classification prob- lems based on CNN and XGBoost

    Thongsuwan S, Jaiyen S, Padcharoen A, Agarwal P . ConvXGB: A new deep learning model for classification prob- lems based on CNN and XGBoost. Nuclear Engineering and T echnology 2021 Feb; https://www.scinapse.io/papers/ 3046626379

    2021

  24. [32]

    Learning to localize temporal events in large-scale video data

    Bober-Irizar M, Skalic M, Austin D. Learning to localize temporal events in large-scale video data. arXiv:191011631 [cs] 2019 Oct;http://arxiv.org/abs/1910.11631, arXiv: 1910.11631

    2019 arXiv

  25. [33]

    The OpenCV Library

    Bradski G. The OpenCV Library. Dr Dobb’s Journal of Software T ools 2000;25(11):120–126

    2000

  26. [34]

    https://zenodo.org/record/6575970, accessed March 30, 2023

    Collette Aea, h5py/h5py: 3.1.0; 2020. https://zenodo.org/record/6575970, accessed March 30, 2023. Zenodo

    2020

  27. [35]

    https://github.com/aleju/imgaug, accessed March 30, 2023

    Jung AB, imgaug: Image augmentation for machine learning experiments; 2017. https://github.com/aleju/imgaug, accessed March 30, 2023. GitHub

    2017

  28. [36]

    T ensorFlow: A system for large-scale machine learning

    Abadi M, Barham P , Chen J, Chen Z, Davis A, Dean J, et al. T ensorFlow: A system for large-scale machine learning. In: Proceedings of the 12th USENIX conference on Operating Systems Design and Implementation OSDI’16, Savannah, GA, USA: USENIX Association; 2016. p. 265–283

    2016

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.