Pith. sign in

REVIEW 3 major objections 6 minor 27 references

Support Vector Machine for Person Classification Using the EEG Signals

T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A multiclass support vector machine trained on principal components of EEG features can attribute a brain-signal window to the correct person among 12 with up to 92.9% accuracy.

desk verdict The 92.9% accuracy is likely inflated by overlapping-window leakage and test-set hyperparameter tuning, so the central result does not survive scrutiny, though the paper is honestly written and cites prior work fairly. read the letter →

arxiv 2411.17446 v1 pith:USETRHPU submitted 2024-11-26 cs.CR eess.SP

Naveenkumar G Venkataswamy , Masudul H Imtiaz This is my paper · ORCID
classification cs.CReess.SP
keywords EEGbiometricspersonidentificationsupportvectormachineprincipalcomponentanalysisuserauthenticationbrain-computerinterfacelivenessdetectiontime-frequencyfeatures
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 asks whether the electrical activity of the brain, recorded with an eight-channel consumer headset, is distinctive enough to serve as a biometric password. It claims the answer is yes: a multiclass support vector machine that reads 27 principal components extracted from 80 time- and frequency-domain features per EEG window can identify which of 12 people produced a given signal, reaching 92.9% accuracy with an RBF kernel. The intended payoff is an authentication method that is hard to forge because the signal comes from a living brain, not from a token or a memorized secret.

What carries the argument

The mechanism is a supervised multiclass support vector machine (SVM), a classifier that separates labeled examples by finding hyperplanes in feature space. Principal component analysis (PCA) serves as the dimensionality-reduction step: it compresses the 80 per-window features into 27 components that preserve 95% of the variance, which the authors say reduces overfitting. The feature set itself mixes simple statistics (root mean square, standard deviation, skewness, kurtosis), Hjorth parameters (activity, mobility, complexity), Shannon and spectral entropy, and power spectral density. The SVM's RBF kernel is the component that actually carries the identification accuracy, and the authors tune C and gamma to reach their best result.

What would settle it

Train the SVM on windows from the first half of each subject's five-minute recording and test on the second half, or on a separate recording session of the same subjects. If accuracy drops toward chance (1/12 ≈ 8.3%), the 92.9% figure is an artifact of overlapping-window leakage; if it remains high, identity information is genuinely stable across time.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that person identity is recoverable from EEG alone. Using a public 12-subject EEG dataset collected for fatigue research, the authors divide each recording into overlapping 0.8-second windows, compute ten features per channel (seven time-domain, three frequency-domain) for 80 total features, reduce these to 27 principal components, and train a multiclass SVM. The RBF kernel with C=100 and gamma=0.01 attains 92.9% accuracy, compared with 81.1% for linear and 84.7% for polynomial kernels. The authors read this as evidence that a practical, spoof-resistant EEG biometric system is within reach, especially if future datasets are collected specifically for authentication.

Load-bearing premise

The central assumption is that randomly splitting overlapping 0.8-second windows from one continuous recording into training and test sets produces independent test samples; because the windows overlap by 0.4 seconds, the test set shares raw signal with the training set, so the accuracy may reflect recording-specific patterns rather than the person's identity.

Editorial extensions

If this is right

  • If the result holds, EEG-based person identification can be done with a low-cost eight-channel headset rather than expensive medical-grade equipment.
  • Because EEG arises from living brain activity, the method could detect liveness and resist spoofing attacks that fool fingerprints or face recognition.
  • The classifier's accuracy depends strongly on kernel choice and hyperparameters (C, gamma), so practical systems will need per-deployment tuning.
  • Using a dataset not originally collected for authentication still yields high accuracy, suggesting identity information is a robust byproduct of EEG recording.
  • The 27 principal components, not the raw 80 features, carry the discriminative information, so a lightweight feature pipeline is sufficient.

Reading between the lines

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

  • The reported 92.9% is likely optimistic because the single continuous recording per subject is split into overlapping windows before random train/test division; adjacent windows share raw signal, so the classifier may be exploiting within-recording correlation rather than stable identity traits. A leave-one-session-out or first-half/second-half split would give a fairer estimate.
  • A practical authentication system would need to verify that accuracy survives across days, changes in electrode placement, and different mental states, since the paper's data comes from one recording session per person.
  • The feature and PCA pipeline could be tested on other EEG datasets with multiple sessions to see whether the 27 principal components generalize or are dataset-specific.
  • The comparison with earlier CNN and SVM results suggests that accuracy is more dependent on evaluation protocol than on classifier family; reporting a common protocol would help the community tell signal from leakage.
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 paper proposes an EEG-based person identification system using multiclass support vector machines. It uses a public 12-subject EEG dataset recorded with an eight-channel OpenBCI headset, extracts seven time-domain and three frequency-domain features per channel (80 features per window), reduces dimensionality with PCA to 27 components, and classifies subjects with linear, polynomial, and RBF kernels. The authors report a maximum identification accuracy of 92.9% with the RBF kernel and claim that this demonstrates the viability of EEG-based biometric authentication.

Significance. If the reported accuracy were credible, the work would provide a useful data point for low-cost EEG biometrics using consumer hardware and a publicly available dataset. The paper has some strengths: it uses a public dataset, specifies a concrete feature set from the literature, and compares three SVM kernels. However, the central empirical claim is not supported by the evaluation as described, because the experimental design permits information leakage between training and test windows and because model selection is performed on the test data. As a result, the 92.9% figure cannot be interpreted as an unbiased estimate of person-classification accuracy.

major comments (3)
  1. [II.B.2 and II.B.5] The windowing procedure creates overlapping 0.8-second windows with 0.4-second overlap. With a 250 Hz sampling rate, adjacent windows share 100 raw samples, so consecutive windows are highly correlated versions of the same EEG segment. The classifier implementation then randomly splits 'each individual's samples' into 80% training and 20% testing. Under this split, a given test window has about a 96% chance of having an immediate neighbor in the training set (since each neighbor is in training with probability 0.8, the chance that at least one of the two neighbors is in training is 1 - 0.2^2 = 0.96). The classifier can therefore match overlapping signal fragments rather than learn a generalizable identity template. This leakage directly inflates the reported accuracy and invalidates the central claim of 92.9% person-classification accuracy.
  2. [III] The reported 'best accuracy' is the maximum over a grid of choices made after seeing the test labels. Section III states that the PCA variance threshold was selected because '95% variance provided the best accuracy,' and that for each kernel the regularization parameter C, polynomial degree, and gamma were varied until the best identification rate was found. There is no separate validation set and no nested cross-validation. Selecting model parameters on the same test split makes the reported numbers optimistically biased. An unbiased protocol would require an independent test set used only once, or nested cross-validation for parameter selection.
  3. [II.B.5 and III] The evaluation is based on a single random split of windowed samples, with all 12 subjects present in both training and testing. This does not measure the ability to identify an unseen person, which is the relevant biometric setting; it measures the ability to classify segments from a recording whose other segments were seen during training. Moreover, no error bars, confidence intervals, or multiple-split statistics are reported for the 81.1%, 84.7%, and 92.9% figures in Table I. A leave-one-subject-out evaluation, or at minimum repeated stratified cross-validation with non-overlapping windows and confidence intervals, is necessary to support the paper's conclusions.
minor comments (6)
  1. [Table I] The column header 'Linaer SVM' is a typo and should read 'Linear SVM.'
  2. [IV] The sentence 'we systematically evaluated each parameter of SVM to implicit the user authentication' appears to contain a word error; it should likely read 'to improve' or 'to implement' user authentication.
  3. [II.B.3] The text says 7 time-domain and 3 frequency-domain features were calculated, but the listed ten features (RMS, standard deviation, skewness, kurtosis, Hjorth activity, Hjorth complexity, Hjorth mobility, Shannon entropy, spectral entropy, power spectral density) are not clearly partitioned into 7 and 3. The authors should clarify which features belong to each category.
  4. [II.B.4 and Fig. 3] The caption of Fig. 3 reads 'the variance explained by every 27 principal components'; it should read 'by each of the 27 principal components.'
  5. [II.B.5] The manuscript does not state the number of windows per subject, the class balance, or the total sample size, which are needed to assess the reliability of the reported accuracy.
  6. [II.B.4] The PCA description does not specify whether features were standardized before PCA, how many components were retained for the 80%, 85%, and 90% thresholds, or whether PCA was fit on the training split only; these details are important for reproducibility.

Circularity Check

2 steps flagged · score 6.0 of 10

The reported 92.9% accuracy is a fitted/memorized value rather than an independent prediction: overlapping windows leak test signal into training, and PCA/SVM parameters are selected on the same test split.

  1. other [Section II.B.2 (Windowing) and II.B.5 (Classifier Implementation)]
    "the EEG signal data is segmented into an overlapping square window, with each window containing 0.8 seconds of EEG data with a 0.4-second overlap with future and previous windows ... we divided each individual’s samples into two parts: 80% dedicated to learning and 20% for testing."

    The test windows are cut from the same continuous recording and share half their raw EEG samples with their immediate neighbors. Because the 80/20 train/test split is applied at the window level after this overlapping segmentation, a typical test window has overlapping data in the training set. High accuracy can therefore be obtained by matching the shared signal fragment, so the reported accuracy is partly constructed from training data rather than measuring identity generalization.

  2. fitted input called prediction [Section II.B.4 (Dimensionality Reduction) and Section III (Results and Discussion)]
    "We used principal component analysis (PCA) with 80%, 85%, 90%, and 95% explained variance. 95% variance provided the best accuracy resulting in 27 principal components. ... Each Kernal function has been modified and its parameters have been adapted to the obtained principal components. ... the gamma parameter is varied, which gives an identification rate equal to 84.7%. For best results for RBF kernel, the parameters c is set to 100 and gamma=0.01 and a level equal to 92.9% is obtained."

    The PCA variance threshold and the SVM kernel hyperparameters (c, degree, gamma) are selected by comparing accuracies on the same test partition, and the headline 92.9% is the maximum of that search. This is parameter fitting on the evaluation data, so the claimed prediction accuracy is not an unbiased out-of-sample estimate; it is a selected maximum of a data-dependent model search.

full rationale

This paper does not exhibit definitional circularity or load-bearing self-citation: the EEG dataset, feature choices, and comparison baselines come from external sources, and the SVM pipeline has real independent content. The central problem is that the headline evaluation is not an independent prediction. First, random splitting of overlapping 0.8 s windows makes test windows overlap with training windows, so the classifier can memorize shared raw signal. Second, PCA variance retention and SVM kernel parameters are tuned against the same test split, and the best value is reported. Both issues make the reported 'maximum accuracy' a fitted or memorized number by construction. Because the core claim is exactly that this accuracy demonstrates EEG-based identification, the circularity is partial but substantial.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central accuracy claim depends on six tuned knobs (PCA variance, number of components, SVM hyperparameters, windowing choices) and on untested domain assumptions about EEG stability and feature discriminability. No new entities are introduced, but the tuned parameters mean the reported accuracy is partly a fitted maximum.

free parameters (6)
  • PCA explained variance threshold = 95%
    Chosen because it 'provided the best accuracy' among 80%, 85%, 90%, and 95% (Section II.B.4).
  • Number of principal components = 27
    Determined by the 95% variance threshold; effectively a model choice tuned on data.
  • Linear SVM regularization C = 10
    Varied C gave the best identification rate at C=10 (Section III).
  • Polynomial SVM C, degree, gamma = C=1; degree and gamma varied, exact values not reported
    Parameters adapted to maximize accuracy; best rate 84.7% (Section III).
  • RBF SVM C and gamma = C=100, gamma=0.01
    Best RBF accuracy of 92.9% obtained with these values (Section III).
  • Window length and overlap = 0.8 seconds, 0.4 seconds overlap
    Chosen by hand in Section II.B.2; affects all downstream features and the accuracy result.
assumptions (5)
  • domain assumption EEG signals are unique and stable per individual
    Introduction states signals are underpinned by unique neural pathways and genetic components, citing [6], but this study offers no validation of temporal stability or uniqueness.
  • domain assumption The fatigue-analysis dataset is suitable for person identification
    Section II.A says the dataset 'serves as a robust foundation', yet it was collected for fatigue analysis with one session per subject and no identity-specific design.
  • domain assumption ASR artifact removal with k=15 does not remove identity-relevant signal
    Section II.B.1 applies ASR with k=15, assuming artifacts are cleanly separable from identity-related EEG.
  • domain assumption The ten literature-selected features capture identity
    Section II.B.3 selects features based on [21,22], assuming these are discriminative for person classification.
  • standard math Standard SVM and PCA theory
    The paper relies on standard properties of PCA and SVM without derivation, which is acceptable as background math.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Support Vector Machine for Person Classification Using the EEG Signals." pith.science (2026). https://pith.science/paper/USETRHPU

@misc{pith2026241117446,
  author       = {Pith},
  title        = {Pith review of: Support Vector Machine for Person Classification Using the EEG Signals},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/USETRHPU}},
  note         = {Machine review of arXiv:2411.17446}
}
read the original abstract

User authentication is a pivotal element in security systems. Conventional methods including passwords, personal identification numbers, and identification tags are increasingly vulnerable to cyber-attacks. This paper suggests a paradigm shift towards biometric identification technology that leverages unique physiological or behavioral characteristics for user authenticity verification. Nevertheless, biometric solutions like fingerprints, iris patterns, facial and voice recognition are also susceptible to forgery and deception. We propose using Electroencephalogram (EEG) signals for individual identification to address this challenge. Derived from unique brain activities, these signals offer promising authentication potential and provide a novel means for liveness detection, thereby mitigating spoofing attacks. This study employs a public dataset initially compiled for fatigue analysis, featuring EEG data from 12 subjects recorded via an eight-channel OpenBCI helmet. This dataset extracts salient features from the EEG signals and trains a supervised multiclass Support Vector Machine classifier. Upon evaluation, the classifier model achieves a maximum accuracy of 92.9\%, leveraging ten features from each channel. Collectively, these findings highlight the viability of machine learning in implementing real-world, EEG-based biometric identification systems, thereby advancing user authentication technology.

Figures

Figures reproduced from arXiv: 2411.17446 by the authors.

Figure 1
Figure 1. The design concept for EEG signal acquisition experiments. Bio￾metric data were collected using OpenBCI in the relaxed state (E.C. and E.O.).) and the A.O. task. In the A.O. tasks, frequent and infrequent stimuli are presented randomly in an 80:20 ratio. The data is transferred to the P.C. via Bluetooth with the OpenViBE software [15]. In total, a 120-second auditory stimulus, which consisted of 2 different tones (2… view at source ↗
Figure 2
Figure 2. Grand average representation of the P300 wave across all the participants. Eight traces (one per channel) are presented for the frontal(F), center(C), parietal (P), and occipital(O) electrodes of the left and right hemispheres of the brain. The shaded area represents standard error across traces [15] these segments, choosing the most informative attributes, and constructing a classifier to identify users. 1) Filteri… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 25 canonical work pages

  1. [1]

    Ethical hacking,

    C. C. Palmer, “Ethical hacking,” IBM Syst. J., vol. 40, no. 3, p. 769–780, mar 2001. [Online]. Available: https://doi.org/10.1147/sj.403.0769

    work page doi:10.1147/sj.403.0769 2001

  2. [2]

    A multimodal biometric system using fingerprint, face and speech,

    A. K. Jain, L. Hong, and Y . Kulkarni, “A multimodal biometric system using fingerprint, face and speech,” in 2nd Int’l Conf. AVBPA , vol. 10, 1999

    work page 1999

  3. [3]

    Biometrics evaluation under spoofing attacks,

    I. Chingovska, A. Anjos, and S. Marcel, “Biometrics evaluation under spoofing attacks,” IEEE Transactions on Information Forensics and Security , vol. 9, pp. 2264–2276, 2014. [Online]. Available: https://api.semanticscholar.org/CorpusID:15501332

    work page 2014

  4. [4]

    Eeg signal analysis: a survey,

    D. P. Subha, P. K. Joseph, R. Acharya U, and C. M. Lim, “Eeg signal analysis: a survey,” Journal of medical systems , vol. 34, pp. 195–212, 2010

    work page 2010

  5. [5]

    Embedded eeg signal acquisition systems,

    L. Losonczi, L. F. M ´arton, T. S. Brassai, and L. Farkas, “Embedded eeg signal acquisition systems,” Procedia Technology, vol. 12, pp. 141–147, 2014

    work page 2014

  6. [6]

    Biometrics from brain electrical activity: A machine learning approach,

    R. Palaniappan and D. P. Mandic, “Biometrics from brain electrical activity: A machine learning approach,” IEEE transactions on pattern analysis and machine intelligence , vol. 29, no. 4, pp. 738–742, 2007

    work page 2007

  7. [7]

    Emotion estimation from eeg signals during listening to quran using psd features,

    M. Alsolamy and A. Fattouh, “Emotion estimation from eeg signals during listening to quran using psd features,” in 2016 7th International Conference on computer science and information technology (CSIT) . IEEE, 2016, pp. 1–5

    work page 2016

  8. [8]

    Survey of eeg-based biometric authentication,

    I. Jayarathne, M. Cohen, and S. Amarakeerthi, “Survey of eeg-based biometric authentication,” in 2017 IEEE 8th International Conference on Awareness Science and Technology (iCAST). IEEE, 2017, pp. 324– 329

    work page 2017

Show all 27 references
  1. [9]

    Human identification from brain eeg signals using advanced machine learning method eeg-based biometrics,

    M. K. Bashar, I. Chiaki, and H. Yoshida, “Human identification from brain eeg signals using advanced machine learning method eeg-based biometrics,” in 2016 IEEE EMBS Conference on Biomedical Engineer- ing and Sciences (IECBES) . IEEE, 2016, pp. 475–479

    2016

  2. [10]

    An independent component analysis (ica) based approach for eeg person authentication,

    C. He and J. Wang, “An independent component analysis (ica) based approach for eeg person authentication,” in 2009 3rd International Conference on Bioinformatics and Biomedical Engineering . IEEE, 2009, pp. 1–4

    2009

  3. [11]

    Comparison of different entropies as features for person authentication based on eeg signals,

    Z. Mu, J. Hu, J. Min, and J. Yin, “Comparison of different entropies as features for person authentication based on eeg signals,” IET Biometrics, vol. 6, no. 6, pp. 409–417, 2017

    2017

  4. [12]

    Eeg-based user authentication using a convolutional neural network,

    T. Yu, C.-S. Wei, K.-J. Chiang, M. Nakanishi, and T.-P. Jung, “Eeg-based user authentication using a convolutional neural network,” in 2019 9th International IEEE/EMBS Conference on Neural Engineering (NER) . IEEE, 2019, pp. 1011–1014

    2019

  5. [13]

    Eeg- based person authentication method using deep learning with visual stimulation,

    S. Puengdang, S. Tuarob, T. Sattabongkot, and B. Sakboonyarat, “Eeg- based person authentication method using deep learning with visual stimulation,” in 2019 11th International Conference on Knowledge and Smart Technology (KST). IEEE, 2019, pp. 6–10

    2019

  6. [14]

    Eeg and empatica e4 signals - five minute p300 test and fas scores,

    M. A. Ram ´ırez-Moreno, P. Carrillo-Tijerina, and J. d. J. Lozoya-Santos, “Eeg and empatica e4 signals - five minute p300 test and fas scores,”

  7. [15]

    Evaluation of a fast test based on biometric signals to assess mental fatigue at the workplace—a pilot study,

    M. A. Ram ´ırez-Moreno, P. Carrillo-Tijerina, M. O. Candela-Leal, M. Alanis-Espinosa, J. C. Tud ´on-Mart´ınez, A. Roman-Flores, R. A. Ram´ırez-Mendoza, and J. d. J. Lozoya-Santos, “Evaluation of a fast test based on biometric signals to assess mental fatigue at the workplace—a...

    2021

  8. [16]

    Skin sympathetic nerve activity and event-related potentials during auditory oddball paradigms,

    H. Ito, Y . Sugiyama, T. Mano, H. Okada, T. Matsukawa, and S. Iwase, “Skin sympathetic nerve activity and event-related potentials during auditory oddball paradigms,” Journal of the autonomic nervous system , vol. 60, no. 3, pp. 129–135, 1996

    1996

  9. [17]

    What is novel in the novelty oddball paradigm? functional significance of the novelty p3 event-related potential as revealed by independent component analysis,

    S. Debener, S. Makeig, A. Delorme, and A. K. Engel, “What is novel in the novelty oddball paradigm? functional significance of the novelty p3 event-related potential as revealed by independent component analysis,” Cognitive Brain Research, vol. 22, no. 3, pp. 309–321, 2005

    2005

  10. [18]

    Openvibe: An open-source software platform to design, test, and use brain–computer interfaces in real and virtual environments,

    Y . Renard, F. Lotte, G. Gibert, M. Congedo, E. Maby, V . Delannoy, O. Bertrand, and A. L ´ecuyer, “Openvibe: An open-source software platform to design, test, and use brain–computer interfaces in real and virtual environments,” Presence, vol. 19, no. 1, pp. 35–53, 2010

    2010

  11. [19]

    Accessible electroencephalograms (eegs): A comparative review with openbci’s ultracortex mark iv headset,

    A. Aldridge, E. Barnes, C. L. Bethel, D. W. Carruth, M. Kocturova, M. Pleva, and J. Juhar, “Accessible electroencephalograms (eegs): A comparative review with openbci’s ultracortex mark iv headset,” in 2019 29th International Conference Radioelektronika (RADIOELEKTRON- IKA). I...

    2019

  12. [20]

    Eeglab: an open source toolbox for analysis of single-trial eeg dynamics including independent component analysis,

    A. Delorme and S. Makeig, “Eeglab: an open source toolbox for analysis of single-trial eeg dynamics including independent component analysis,” Journal of neuroscience methods , vol. 134, no. 1, pp. 9–21, 2004

    2004

  13. [21]

    Multimodal eeg and keystroke dynamics based biometric system using machine learning algorithms,

    A. Rahman, M. E. Chowdhury, A. Khandakar, S. Kiranyaz, K. S. Zaman, M. B. I. Reaz, M. T. Islam, M. Ezeddin, and M. A. Kadir, “Multimodal eeg and keystroke dynamics based biometric system using machine learning algorithms,” IEEE Access, vol. 9, pp. 94 625–94 643, 2021

    2021

  14. [22]

    Stable eeg features for biometric recognition in resting state conditions,

    D. La Rocca, P. Campisi, and G. Scarano, “Stable eeg features for biometric recognition in resting state conditions,” in Biomedical Engi- neering Systems and Technologies: 6th International Joint Conference, BIOSTEC 2013, Barcelona, Spain, February 11-14, 2013, Revised Selecte...

    2013

  15. [23]

    Seizure lateralization in scalp eeg using hjorth parameters t. cecchin, r. ranta b, o. caspary c, l. koessler b, d, h. vespignani b, d, l. maillard b, d

    T. Cecchin, “Seizure lateralization in scalp eeg using hjorth parameters t. cecchin, r. ranta b, o. caspary c, l. koessler b, d, h. vespignani b, d, l. maillard b, d.”

  16. [24]

    Feature extraction of eeg signals using power spectral entropy,

    A. Zhang, B. Yang, and L. Huang, “Feature extraction of eeg signals using power spectral entropy,” in 2008 international conference on BioMedical engineering and informatics , vol. 2. IEEE, 2008, pp. 435– 439

    2008

  17. [25]

    A comparison between power spectral density and network metrics: an eeg study,

    M. Demuru, S. M. La Cava, S. M. Pani, and M. Fraschini, “A comparison between power spectral density and network metrics: an eeg study,” Biomedical Signal Processing and Control , vol. 57, p. 101760, 2020

    2020

  18. [26]

    A meta-analysis of overfitting in machine learning,

    R. Roelofs, V . Shankar, B. Recht, S. Fridovich-Keil, M. Hardt, J. Miller, and L. Schmidt, “A meta-analysis of overfitting in machine learning,” Advances in Neural Information Processing Systems , vol. 32, 2019

    2019

  19. [2021]

    Available: https://dx.doi.org/10.21227/0sjm-r993

    [Online]. Available: https://dx.doi.org/10.21227/0sjm-r993

    doi

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.