Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

CRIA: A Cross-View Interaction and Instance-Adapted Pre-training Framework for Generalizable EEG Representations

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read CRIA claims that treating EEG as three interacting views, with the spectral view as anchor, yields generalizable representations that beat matched pretraining baselines on TUAB, TUEV, and transfer to CHB-MIT.

desk verdict Solid empirical EEG pretraining paper with a real cross-dataset transfer claim, undermined by a bogus information-bottleneck derivation and an undocumented channel-mapping step in the CHB-MIT transfer. read the letter →

arxiv 2506.16056 v1 pith:6PWKSQVV submitted 2025-06-19 cs.LG cs.AI

Puchun Liu , C. L. Philip Chen , Yubin He , Tong Zhang This is my paper · ORCID
classification cs.LGcs.AI
keywords EEGrepresentationlearningself-supervisedpretrainingcross-viewcross-attentionvariable-lengthencodinginformationbottleneckeventclassificationanomalydetection
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

CRIA is a self-supervised pretraining method that reads an EEG recording as three parallel views rather than one flat token sequence: temporal, spectral, and spatial. The paper argues that these views carry complementary information, and that making the spectral view the stable anchor while temporal and spatial features cross-attend to it lets a pretrained encoder learn representations that generalize across datasets and tasks. On the TUH EEG corpus, CRIA reports 80.03% balanced accuracy for abnormal-EEG detection and 57.02% for multi-class event classification; after pretraining on TUH, it reports 83.79% balanced accuracy on the unseen CHB-MIT dataset. The paper reports that all models were evaluated on the same data splits, and the main large-scale baseline was re-pretrained under matched data and scale, so the claimed gains are attributed to the multi-view interaction and the view-wise masking objective rather than to larger pretraining corpora. A reader should care because EEG datasets differ in sampling rate, channel count, and montage, and CRIA is designed to absorb that heterogeneity instead of requiring every dataset to be reformatted into one fixed template.

What carries the argument

The load-bearing object is an asymmetric three-view cross-attention block. In each layer the spectral view is refined by self-attention, while the temporal and spatial views are updated by cross-attending to it: temporal and spatial features supply the queries $Q$, and spectral features supply the keys $K$ and values $V$. Around this block sit four supporting mechanisms: instance-adaptive encoding, which uses a learnable channel embedding $E_{\mathrm{channel}}\in\mathbb{R}^{C_{\max}\times D}$, rotary position encoding, and FFT-derived spectral features so one encoder can consume variable segment lengths and channel counts; a view-wise masking pretraining task in which one view is replaced by a learnable placeholder and a contrastive loss (temperature $T=0.2$) pulls the masked representation toward the unmasked one; a representation-enhanced purification step that selects top-$k_c$ channels and top-$k_t$ segments before pooling so salient event regions dominate the fused feature; and randomized attention-value masking during fine-tuning, justified by an information-bottleneck argument that masking reduces $I(A;X)$. The mechanism's job is to exchange complementary information across views while keeping the spectral view stable enough to anchor learning on non-stationary, low-signal-to-noise EEG.

What would settle it

Pretrain on one montage and fine-tune on a second montage that has the same channel names but different electrode positions; if cross-montage transfer falls to the level of training from scratch while within-montage transfer stays high, the shared channel-embedding assumption is the failure point. A second check: use a downstream task whose events span several tokens or start mid-token; if event-aligned tokenization beats the fixed-length version substantially, the segmentation is losing information.

Watch

Extended reading notes

Core claim

The central discovery CRIA argues for is that EEG self-supervised learning benefits from modeling views as asymmetric participants: one view (spectral, obtained by FFT amplitude) is held as the stable reference, and the temporal and spatial views are used as queries in cross-attention whose keys and values come from the spectral view. This is paired with a view-wise masking pretraining task in which one of the three views is randomly replaced by a learnable placeholder and the model is trained, through a contrastive loss, to keep its representation consistent with the unmasked branch. The architecture also uses a learnable channel embedding over the union of all channel labels and linear attention with rotary position encoding, so a single pretrained encoder accepts variable segment lengths and channel counts. The paper reports that this combination outperforms baselines with matched pretraining data on TUAB and TUEV, transfers to CHB-MIT better than previous pretrained EEG models, retains an advantage when trained from scratch, and degrades less under injected noise; it also draws attention to "cross-view feature emergence," where salient patterns from one view appear in another view in later layers.

Load-bearing premise

Channels that share a label are treated as the same brain location across different montages and acquisition systems, and slicing any recording into fixed-length tokens is assumed to preserve the events that downstream tasks need.

Editorial extensions

If this is right

  • If the reported numbers hold, EEG pretraining no longer needs to force every dataset into a fixed channel template: a single encoder pretrained on a union of channel labels can be fine-tuned directly on recordings with different montages and lengths.
  • The view-wise masking task is an easier pretraining objective than full reconstruction or perturbation, so pretraining should converge in fewer steps and stay effective on modest or noisy pretraining corpora.
  • The asymmetric design makes a concrete prediction: removing the spectral view should hurt downstream performance more than removing either of the other two views, and corrupting the spectral input during transfer should be the most damaging noise.
  • The from-scratch and noise-resilience results suggest the cross-view architecture itself, not just the pretraining data, is doing much of the representational work, so the design may transfer to other non-stationary biosignals.

Reading between the lines

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

  • A testable extension the paper leaves implicit is to evaluate transfer to a target montage that shares no channel labels with the pretraining montage; the reported CHB-MIT transfer probably depends on the shared channel vocabulary, and a fallback embedding would be needed for zero-overlap montages.
  • The information-bottleneck justification works for any random mask independent of the data, so the same argument licenses per-layer attention-dropout rates; tuning the mask fraction per layer, rather than using one fixed value, could push the noise-resilience numbers further.
  • If the observed cross-view feature emergence is genuine, the spectral-anchor attention maps could double as a saliency tool for clinicians, linking learned cross-view patterns back to the event-related channels and time windows a neurologist would mark.
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 / 4 minor

Summary. The paper proposes CRIA, a self-supervised pre-training framework for EEG that combines temporal, spectral, and spatial views through an asymmetric cross-attention architecture anchored on the spectral view. It introduces an instance-adaptive variable-length/variable-channel encoding, a view-wise masking contrastive pre-training objective, and a randomized attention-matrix mask during fine-tuning that the authors claim is justified by information-bottleneck theory. Experiments are carried out on the TUH corpus (TUAB, TUEV) and cross-dataset transfer to CHB-MIT, with comparisons to several baselines including a same-scale re-pretrained LaBraM, along with ablations and noise-robustness studies. The paper reports state-of-the-art balanced accuracy on TUAB (80.03%) and TUEV (57.02%), and strong transfer results on CHB-MIT.

Significance. If the empirical results are reproducible, CRIA is a useful contribution to EEG representation learning: it addresses multi-view interaction, variable channel/length heterogeneity, and cross-dataset transfer, and the experimental protocol is more controlled than many prior efforts because LaBraM is re-pretrained on the same data and scale. The paper's strengths include five-seed reporting, public datasets, detailed ablations of the view-fusion and purification components, and a clear statement that code will be released. However, the theoretical justification of the attention-masking strategy contains a mathematical error, and the cross-dataset transfer to CHB-MIT rests on an undocumented channel-identity mapping. These two issues affect load-bearing claims in the paper (the IB-based validation and the strong generalization/tranfer conclusion), so the manuscript needs substantial revision before the claims can be accepted as stated.

major comments (3)
  1. [Section 3.4, Eqs. (23)–(24)] The derivation that the randomized attention mask is supported by the information bottleneck contains a false step. From Eq. (24) the authors correctly obtain I(A;X|M) = I(A';X) + I(M;X|A') and therefore I(A;X|M) ≥ I(A';X). However, the next sentence, "According to the mathematical definition of information, it follows that I(A;X) ≥ I(A;X|M)," is not true: conditioning can increase mutual information. In fact, because M is assumed independent of X, I(A;X|M) = I(A;X) + I(M;X|A) ≥ I(A;X), so the inequality is reversed. The chain I(A;X) ≥ I(A;X|M) ≥ I(A';X) does not follow, and the claim that this strategy is "validated through information bottleneck theory" (as stated in the contributions) is unsupported. Please correct the derivation or remove/replace the IB-based justification; the empirical effect of the mask can stand on its own.
  2. [Section 3.2 / 3.5 / Table 2] The cross-dataset transfer to CHB-MIT relies on an undocumented channel-identity mapping. E_channel is defined as a learnable embedding of dimension C_max, where C_max is the maximum number of channels in the pre-training dataset (Section 3.2, Eq. (7)–(9)). Pre-training is performed on the TUH Corpus only (Section 3.5), while CHB-MIT is a different acquisition with montage-specific bipolar channel labels (e.g., FP1-F7). The paper never describes how unseen CHB-MIT channel labels are mapped to entries of E_channel. If the implementation indexes channels by position or uses a separate ad-hoc mapping, then Table 2 does not evaluate the pretrained encoder's cross-dataset channel generalization but rather an undocumented adaptation. This is load-bearing because the "strong generalization ability" conclusion is based in part on the CHB-MIT results. Please disclose the exact mapping, or treat CHB-MIT channel embeddings as separately trained and show that the reported transfer gains are robust to that choice.
  3. [Table 2 and Abstract] The statement that CRIA "outperforms existing methods" on cross-dataset transfer is not supported by the PR AUC column of Table 2: CRIA obtains PR AUC 0.3982±.0164, which is lower than LaBraM's 0.4007±.0150. The text in Section 4.2 claims the model "far surpasses" other models on CHB-MIT, but the advantage is only in BACC and AUROC, not in PR AUC. Since the abstract claims broad superiority, please either qualify the claim (e.g., "higher balanced accuracy and AUROC, comparable PR AUC") or provide a statistical comparison that justifies the stated conclusion despite the lower PR AUC.
minor comments (4)
  1. [Eq. (18)–(20)] The notation for the purification strategy is confusing: C_kc is defined via a topk operation but then used as a set of channels in Eq. (20); T_kt is similarly used as a set. Please clarify with set-builder notation and make explicit that the top-k selection is by L2 norm over the appropriate axis.
  2. [Fig. 5] The ablation bar chart reports only BACC and Cohen's Kappa; please provide error bars or significance indicators for the ablations, since the text states that all experiments passed a t-test but the figure does not show variability.
  3. [Section 3.3.1, Eq. (15)] In Eq. (15), CA(Q_{v1}, K_{v2}, V_{v2}) = softmax(Q_{v1}K_{v2}^T / sqrt(d_k)) V_{v2}; the notation should be consistent with the self-attention equation, and the role of the scaling factor d_k should be defined clearly (it is already defined for self-attention but not repeated here).
  4. [Section 4.1, Table 1] BrainBERT and EEGFORMER rows report only AUROC and PR AUC for TUAB but no TUEV results; the text says "—" means not mentioned, but readers may benefit from a footnote explaining whether these models were not run on TUEV or were run but omitted.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CRIA's empirical claims rest on held-out downstream evaluation after TUH pretraining, not on fitted constants, self-citation chains, or definitional equivalences.

full rationale

The paper's central claims are empirical and self-contained. CRIA is pretrained on unlabeled TUH Corpus data and then evaluated on TUAB, TUEV, and CHB-MIT using held-out test splits with multiple random seeds; the reported balanced accuracies are downstream classification results, not quantities reconstructed from pretraining objectives or fitted parameters. The channel embedding E_channel is a learned component of the model, and no experiment claims to predict a number that was used to define or fit that embedding. The attention-masking justification via information bottleneck is a derivation from stated assumptions, and although it may be mathematically questionable, it is not circular: the conclusion that masking reduces I(A;X) is not assumed as an input. The CHB-MIT transfer experiment raises a potential validity concern about undocumented channel-label alignment across montages, but that is a correctness and reporting issue, not a circularity where the prediction reduces to the input by construction. Self-citations appear in related work and routine methodological references, but none is load-bearing for the paper's central derivation or empirical claims; no uniqueness theorem or pre-commitment from the authors' prior work is used to force the model design. Overall, the derivation chain from pretraining objective to downstream performance does not collapse into its own inputs.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on several unproven domain assumptions: EEG preprocessing choices (200 Hz resampling, 0.5 to 120 Hz filter) preserve task-relevant information; one shared channel embedding aligns different montages; the spectral view is a stable anchor; and the attention masking benefit follows from a mutual-information inequality that is false as stated. The model also depends on tunable constants (T=0.2, top-k fractions, masking ratio, layer count) whose values are not fully disclosed.

free parameters (4)
  • contrastive temperature T = 0.2
    Set globally in Section 3.4 as 'T=0.2 throughout the paper'; no sensitivity analysis is reported, and downstream results depend on it.
  • top-k fractions k_c and k_t = not reported
    Used in representation purification (Eqs. 18 to 19); values are selected on validation data but not stated in the main text.
  • random attention masking ratio = not reported
    A 'certain percentage' of attention values is masked during fine-tuning in Section 3.4; the percentage is not disclosed and is chosen by hand.
  • per-dataset fine-tuning layer count = 5 pretrain layers; full layers for TUAB and CHB-MIT; last 3 layers for TUEV
    Layer selection differs by downstream task and is chosen on validation, which can favor the reported metrics.
assumptions (6)
  • ad hoc to paper Conditioning reduces mutual information: I(A;X) >= I(A;X|M).
    Invoked in Section 3.4 to connect attention masking to information bottleneck; this inequality is false in general and is the pivotal step in Eqs. 23 to 24.
  • domain assumption Shared learnable channel embedding E_channel over the union of electrode labels makes different montages comparable.
    Section 3.2 states one E_channel over Cmax is used for all datasets; cross-dataset transfer to CHB-MIT depends on this alignment.
  • domain assumption Resampling all EEG to 200 Hz and filtering to 0.5 to 120 Hz preserves task-relevant information across datasets.
    Section 3.1 sets these preprocessing choices for semantic alignment, following clinical EEG practice; if high-frequency or very slow components carry task signal, this loses it.
  • domain assumption The spectral view is stable and low-noise, so FFT amplitude features can serve as the dominant Key and Value anchor.
    Section 3.3.1 designates the spectral view as primary to avoid training instability; this is asserted, not demonstrated across datasets.
  • standard math Standard RoPE, Linformer linear attention, and FFT operate correctly on variable-length segment inputs.
    Used in Section 3.2 without formal proof; these are standard results from the cited literature.
  • ad hoc to paper A random view mask creates an easier pretraining task that improves downstream convergence.
    Section 4.6 compares it to BIOT perturbation on one dataset; the claim is empirical and not supported by a theory.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CRIA: A Cross-View Interaction and Instance-Adapted Pre-training Framework for Generalizable EEG Representations." pith.science (2026). https://pith.science/paper/6PWKSQVV

@misc{pith2026250616056,
  author       = {Pith},
  title        = {Pith review of: CRIA: A Cross-View Interaction and Instance-Adapted Pre-training Framework for Generalizable EEG Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6PWKSQVV}},
  note         = {Machine review of arXiv:2506.16056}
}
read the original abstract

The difficulty of extracting deep features from EEG data and effectively integrating information from multiple views presents significant challenges for developing a generalizable pretraining framework for EEG representation learning. However, most existing pre-training methods rely solely on the contextual semantics of a single view, failing to capture the complex and synergistic interactions among different perspectives, limiting the expressiveness and generalization of learned representations. To address these issues, this paper proposes CRIA, an adaptive framework that utilizes variable-length and variable-channel coding to achieve a unified representation of EEG data across different datasets. In this work, we define cross-view information as the integrated representation that emerges from the interaction among temporal, spectral, and spatial views of EEG signals. The model employs a cross-attention mechanism to fuse temporal, spectral, and spatial features effectively, and combines an attention matrix masking strategy based on the information bottleneck principle with a novel viewpoint masking pre-training scheme. Experimental results on the Temple University EEG corpus and the CHB-MIT dataset show that CRIA outperforms existing methods with the same pre-training conditions, achieving a balanced accuracy of 57.02% for multi-class event classification and 80.03% for anomaly detection, highlighting its strong generalization ability.

Figures

Figures reproduced from arXiv: 2506.16056 by the authors.

Figure 1
Figure 1. CRIA’s overall framework structure. The top box represents the downstream fine-tuning, and the [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Feature visualizations of frequency and spatial views across di [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Spatiotemporal regions exhibiting significant fluctuations in a given EEG event labeled by a human [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Flowchart of pre-training and downstream fine-tuning. [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]
Figure 5
Figure 5. Figure 5: Ablation validation for CRIA in several strategies on TUEV. [PITH_FULL_IMAGE:figures/full_fig_p023_5.png]
Figure 6
Figure 6. Figure 6: Model Robustness Experimental Results. Adding di [PITH_FULL_IMAGE:figures/full_fig_p025_6.png]
Figure 7
Figure 7. Figure 7: Changes in downstream performance of Perturb and Maskview strategies as the number of pre [PITH_FULL_IMAGE:figures/full_fig_p028_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Large-Scale AI and Foundation Models for Neuroscience: A Comprehensive Review

    cs.AI 2025-10 conditional novelty 1.0 of 10

    This paper is a survey: it organizes existing foundation-model work in neuroscience into five application domains and lists public datasets, without presenting new experiments.

Reference graph

Works this paper leans on

53 extracted references · 37 canonical work pages · cited by 1 Pith paper

  1. [1]

    J. Jing, W. Ge, S. Hong, M. B. Fernandes, Z. Lin, C. Yang, S. An, A. F. Struck, A. Herlopian, I. Karakis, et al., Development of expert-level classification of seizures and rhythmic and periodic patterns during eeg interpretation, Neurology 100 (17) (2023) e1750–e1762

    work page 2023

  2. [2]

    Arroyo, S

    S. Arroyo, S. Uematsu, High-frequency eeg activity at the start of seizures, Jour- nal of Clinical Neurophysiology 9 (3) (1992) 441–448

    work page 1992

  3. [3]

    Taran, P

    S. Taran, P. C. Sharma, V . Bajaj, Automatic sleep stages classification using opti- mize flexible analytic wavelet transform, Knowledge-Based Systems 192 (2020) 105367

    work page 2020

  4. [4]

    Z. Jia, Y . Lin, J. Wang, R. Zhou, X. Ning, Y . He, Y . Zhao, Graphsleepnet: Adap- tive spatial-temporal graph convolutional networks for sleep stage classification., in: Ijcai, V ol. 2021, 2020, pp. 1324–1330

    work page 2021

  5. [5]

    Zhang, F

    J. Zhang, F. Zhang, H. Wei, Psss-eeg: A probabilistic-masking self-supervised swin-transformer model for eeg-based drowsiness recognition, Pattern Recogni- tion 158 (2025) 111005

    work page 2025

  6. [6]

    X. Tang, Y . Xie, X. Li, B. Wang, Riding feeling recognition based on multi- head self-attention lstm for driverless automobile, Pattern Recognition 159 (2025) 111135. 29

    work page 2025

  7. [7]

    Murugappan, S

    M. Murugappan, S. Murugappan, Human emotion recognition through short time electroencephalogram (eeg) signals using fast fourier transform (FFT), in: 2013 IEEE 9th International Colloquium on Signal Processing and its Applications, IEEE, 2013, pp. 289–294

    work page 2013

  8. [8]

    T. Song, W. Zheng, P. Song, Z. Cui, Eeg emotion recognition using dynamical graph convolutional neural networks, IEEE Transactions on Affective Computing 11 (3) (2018) 532–541

    work page 2018

Show all 53 references
  1. [9]

    Zhang, X

    T. Zhang, X. Wang, X. Xu, C. P. Chen, Gcb-net: Graph convolutional broad net- work and its application in emotion recognition, IEEE Transactions on Affective Computing 13 (1) (2019) 379–388

    2019

  2. [10]

    Zhang, C

    S. Zhang, C. Tang, C. Guan, Visual-to-eeg cross-modal knowledge distillation for continuous emotion recognition, Pattern Recognition 130 (2022) 108833

    2022

  3. [11]

    Werkle-Bergner, V

    M. Werkle-Bergner, V . Müller, S.-C. Li, U. Lindenberger, Cortical eeg correlates of successful memory encoding: implications for lifespan comparisons, Neuro- science & Biobehavioral Reviews 30 (6) (2006) 839–854

    2006

  4. [12]

    Devlin, Bert: Pre-training of deep bidirectional transformers for language un- derstanding, arXiv preprint arXiv:1810.04805 (2018)

    J. Devlin, Bert: Pre-training of deep bidirectional transformers for language un- derstanding, arXiv preprint arXiv:1810.04805 (2018)

    2018 arXiv

  5. [13]

    Achiam, S

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, et al., Gpt-4 technical report, arXiv preprint arXiv:2303.08774 (2023)

    2023 arXiv

  6. [14]

    C. Wang, V . Subramaniam, A. U. Yaari, G. Kreiman, B. Katz, I. Cases, A. Barbu, Brainbert: Self-supervised representation learning for intracranial recordings, arXiv preprint arXiv:2302.14367 (2023)

    2023 arXiv

  7. [15]

    C. Yang, M. Westover, J. Sun, Biot: Biosignal transformer for cross-data learning in the wild, Advances in Neural Information Processing Systems 36 (2024)

    2024

  8. [16]

    Zhang, Z

    D. Zhang, Z. Yuan, Y . Yang, J. Chen, J. Wang, Y . Li, Brant: Foundation model for intracranial neural signal, Advances in Neural Information Processing Systems 36 (2024). 30

    2024

  9. [17]

    Z. Yuan, D. Zhang, J. Chen, G. Gu, Y . Yang, Brant-2: Foundation model for brain signals, arXiv preprint arXiv:2402.10251 (2024)

    2024 arXiv

  10. [18]

    Jiang, L.-M

    W.-B. Jiang, L.-M. Zhao, B.-L. Lu, Large brain model for learning generic rep- resentations with tremendous eeg data in bci, arXiv preprint arXiv:2405.18765 (2024)

    2024 arXiv

  11. [19]

    S. S. Bhatti, A. Yadav, M. Monga, N. Kumar, Comparative analysis of deep learning approaches for harmful brain activity detection using eeg, arXiv preprint arXiv:2412.07878 (2024)

    2024

  12. [20]

    F. L. da Silva, Eeg and meg: relevance to neuroscience, Neuron 80 (5) (2013) 1112–1128

    2013

  13. [21]

    Allen, Short term spectral analysis, synthesis, and modification by discrete fourier transform, IEEE transactions on acoustics, speech, and signal processing 25 (3) (1977) 235–238

    J. Allen, Short term spectral analysis, synthesis, and modification by discrete fourier transform, IEEE transactions on acoustics, speech, and signal processing 25 (3) (1977) 235–238

    1977

  14. [22]

    Daubechies, The wavelet transform, time-frequency localization and signal analysis, IEEE transactions on information theory 36 (5) (1990) 961–1005

    I. Daubechies, The wavelet transform, time-frequency localization and signal analysis, IEEE transactions on information theory 36 (5) (1990) 961–1005

    1990

  15. [23]

    F. A. Alturki, K. AlSharabi, A. M. Abdurraqeeb, M. Aljalal, Eeg signal anal- ysis for diagnosing neurological disorders using discrete wavelet transform and intelligent techniques, Sensors 20 (9) (2020) 2505

    2020

  16. [24]

    Tabejamaat, H

    M. Tabejamaat, H. Mohammadzade, F. Negin, F. Bremond, Eeg classification with limited data: A deep clustering approach, Pattern Recognition 157 (2025) 110934

    2025

  17. [25]

    R. T. Schirrmeister, J. T. Springenberg, L. D. J. Fiederer, M. Glasstetter, K. Eggensperger, M. Tangermann, F. Hutter, W. Burgard, T. Ball, Deep learning with convolutional neural networks for eeg decoding and visualization, Human brain mapping 38 (11) (2017) 5391–5420

    2017

  18. [26]

    Zheng, W

    X. Zheng, W. Chen, Y . You, Y . Jiang, M. Li, T. Zhang, Ensemble deep learning for automated visual classification using eeg signals, Pattern Recognition 102 (2020) 107147. 31

    2020

  19. [27]

    Craik, Y

    A. Craik, Y . He, J. L. Contreras-Vidal, Deep learning for electroencephalogram (eeg) classification tasks: a review, Journal of neural engineering 16 (3) (2019) 031001

    2019

  20. [28]

    Y . Ding, N. Robinson, C. Tong, Q. Zeng, C. Guan, Lggnet: Learning from local- global-graph representations for brain–computer interface, IEEE Transactions on Neural Networks and Learning Systems (2023)

    2023

  21. [29]

    C.-F. R. Chen, Q. Fan, R. Panda, Crossvit: Cross-attention multi-scale vision transformer for image classification, in: Proceedings of the IEEE/CVF interna- tional conference on computer vision, 2021, pp. 357–366

    2021

  22. [30]

    Pahuja, S

    S. Pahuja, S. Cai, T. Schultz, H. Li, Xanet: cross-attention between eeg of left and right brain for auditory attention decoding, in: 2023 11th International IEEE/EMBS Conference on Neural Engineering (NER), IEEE, 2023, pp. 1–4

    2023

  23. [31]

    P. Gao, T. Liu, J.-W. Liu, B.-L. Lu, W.-L. Zheng, Multimodal multi-view spectral- spatial-temporal masked autoencoder for self-supervised emotion recognition, in: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2024, pp...

    2024

  24. [32]

    X. Wei, K. Zhao, Y . Jiao, N. B. Carlisle, H. Xie, G. A. Fonzo, Y . Zhang, Multi- modal cross-domain self-supervised pre-training for fmri and eeg fusion, Neural Networks 184 (2025) 107066

    2025

  25. [33]

    J. Wang, S. Zhao, Z. Luo, Y . Zhou, H. Jiang, S. Li, T. Li, G. Pan, Cbramod: A criss-cross brain foundation model for eeg decoding, arXiv preprint arXiv:2412.07236 (2024)

    2024

  26. [34]

    Gotman, High frequency oscillations: the new eeg frontier?, Epilepsia 51 (Suppl 1) (2010) 63

    J. Gotman, High frequency oscillations: the new eeg frontier?, Epilepsia 51 (Suppl 1) (2010) 63

    2010

  27. [35]

    R. J. Staba, C. L. Wilson, A. Bragin, D. Jhung, I. Fried, J. Engel Jr, High- frequency oscillations recorded in human medial temporal lobe during sleep, An- nals of neurology 56 (1) (2004) 108–115. 32

    2004

  28. [36]

    S. Wang, B. Z. Li, M. Khabsa, H. Fang, H. Ma, Linformer: Self-attention with linear complexity, arXiv preprint arXiv:2006.04768 (2020)

    2020 arXiv

  29. [37]

    J. Su, M. Ahmed, Y . Lu, S. Pan, W. Bo, Y . Liu, Roformer: Enhanced transformer with rotary position embedding, Neurocomputing 568 (2024) 127063

    2024

  30. [38]

    Y . Chen, K. Ren, K. Song, Y . Wang, Y . Wang, D. Li, L. Qiu, Eegformer: Towards transferable and interpretable large-scale eeg foundation model, arXiv preprint arXiv:2401.10278 (2024)

    2024 arXiv

  31. [39]

    K. He, H. Fan, Y . Wu, S. Xie, R. Girshick, Momentum contrast for unsupervised visual representation learning, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 9729–9738

    2020

  32. [40]

    A. M. Saxe, Y . Bansal, J. Dapello, M. Advani, A. Kolchinsky, B. D. Tracey, D. D. Cox, On the information bottleneck theory of deep learning, Journal of Statistical Mechanics: Theory and Experiment 2019 (12) (2019) 124020

    2019

  33. [41]

    Clevert, Fast and accurate deep network learning by exponential linear units (elus), arXiv preprint arXiv:1511.07289 (2015)

    D.-A. Clevert, Fast and accurate deep network learning by exponential linear units (elus), arXiv preprint arXiv:1511.07289 (2015)

    2015 arXiv

  34. [42]

    Lin, Focal loss for dense object detection, arXiv preprint arXiv:1708.02002 (2017)

    T. Lin, Focal loss for dense object detection, arXiv preprint arXiv:1708.02002 (2017)

    2017 arXiv

  35. [43]

    Zhang, M

    Z. Zhang, M. Sabuncu, Generalized cross entropy loss for training deep neural networks with noisy labels, Advances in neural information processing systems 31 (2018)

    2018

  36. [44]

    Harati, S

    A. Harati, S. López, I. Obeid, J. Picone, M. P. Jacobson, S. Tobochnik, The tuh eeg corpus: A big data resource for automated eeg interpretation, in: 2014 IEEE Signal Processing in Medicine and Biology Symposium (SPMB), 2014, pp. 1–5. doi:10.1109/SPMB.2014.7002953

    2014

  37. [45]

    Guttag, Chb-mit scalp eeg database (version 1.0

    J. Guttag, Chb-mit scalp eeg database (version 1.0. 0), physionet,(2010), URL: https://doi. org/10.13026/C2K01R. 33

    2010 doi

  38. [46]

    A. H. Shoeb, Application of machine learning to epileptic seizure onset detection and treatment, Ph.D. thesis, Massachusetts Institute of Technology (2009)

    2009

  39. [47]

    A. L. Goldberger, L. A. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, H. E. Stanley, Physiobank, phys- iotoolkit, and physionet: components of a new research resource for complex physiologic signals, circulation 101 (23) (2...

    2000

  40. [48]

    C. Yang, D. Xiao, M. B. Westover, J. Sun, Self-supervised eeg representation learning for automatic sleep staging, arXiv preprint arXiv:2110.15278 (2021)

    2021 arXiv

  41. [49]

    H. Li, M. Ding, R. Zhang, C. Xiu, Motor imagery eeg classification algorithm based on cnn-lstm feature fusion network, Biomedical signal processing and con- trol 72 (2022) 103342

    2022

  42. [50]

    W. Y . Peh, Y . Yao, J. Dauwels, Transformer convolutional neural networks for au- tomated artifact detection in scalp eeg, in: 2022 44th Annual International Con- ference of the IEEE Engineering in Medicine & Biology Society (EMBC), IEEE, 2022, pp. 3599–3602

    2022

  43. [51]

    Y . Song, X. Jia, L. Yang, L. Xie, Transformer-based spatial-temporal feature learning for eeg decoding, arXiv preprint arXiv:2106.11170 (2021)

    2021 arXiv

  44. [52]

    N. M. Foumani, G. Mackellar, S. Ghane, S. Irtza, N. Nguyen, M. Salehi, Eeg2rep: enhancing self-supervised eeg representation through informative masked inputs, arXiv preprint arXiv:2402.17772 (2024)

    2024 arXiv

  45. [53]

    Van Den Oord, O

    A. Van Den Oord, O. Vinyals, et al., Neural discrete representation learning, Ad- vances in neural information processing systems 30 (2017). 34

    2017

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.