Pith. sign in

REVIEW 3 major objections 6 minor 56 references

STEAM ranks first across 14 EEG evaluation settings using a shared soft mixture-of-experts backbone.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 15:47 UTC pith:PCMNI7YD

load-bearing objection A well-built EEG foundation-model paper with a novel architecture, but its headline rank claim is compromised because the model configuration was selected on the same benchmark used for the claim. the 3 major comments →

arxiv 2608.02070 v1 pith:PCMNI7YD submitted 2026-08-03 cs.CV cs.LG

STEAM:ASpatio-TEmporal Alignment Mixture-of-Experts Model with Hierarchical Pre-training for EEG Decoding

Zhu Chen , Dingkun Liu , Yuheng Chen , Dongrui Wu This is my paper
classification cs.CV cs.LG
keywords EEG decodingbrain-computer interfacefoundation modelmixture-of-expertsspatio-temporal representationhierarchical pretrainingtransfer learningparameter-efficient adaptation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to settle a three-way tension in EEG foundation models: general transferability across paradigms, accurate decoding within a paradigm, and cheap downstream adaptation. It presents STEAM, a hierarchical pretraining framework that first learns a reusable spatio-temporal backbone from large unlabeled EEG, then continues pretraining on multiple datasets from a target paradigm with supervised objectives. The backbone is a dual-branch encoder—one branch treats channels as tokens, the other treats time segments as tokens—coupled by a shared soft mixture-of-experts (SSMoE) module with eight soft slots, plus frequency-aware gated attention. On seven downstream datasets under leave-one-subject-out and few-shot protocols, STEAM's Stage-I model attains the best average rank among the methods compared, and the Stage-II specialized variants improve paradigm-specific accuracy further. If correct, this would make STEAM currently the strongest all-round EEG foundation model on this benchmark, with a lightweight adaptation mode that updates only about 5.1% of parameters.

Core claim

On its own terms, the paper's central discovery is that general EEG pretraining and paradigm-specific specialization need not be traded off: a two-stage general-to-paradigm pretraining schedule, applied to a dual-branch spatial/temporal encoder with a shared soft mixture-of-experts interaction, yields a single backbone that transfers broadly and then specializes efficiently. The evidence is the benchmark: across seven datasets and fourteen evaluation settings, STEAM achieves a combined rank of 1.43 against a mean of 10.55 for ten foundation-model baselines, at competitive FLOPs. The shared SSMoE—where spatial and temporal tokens are routed through a compact pool of eight soft slots and redis

What carries the argument

The load-bearing mechanism is the shared soft mixture-of-experts (SSMoE) module. It concatenates the spatial-branch tokens and temporal-branch tokens, computes token-to-slot affinities through a learnable routing matrix, aggregates tokens into a compact set of eight soft slots (one expert per slot), transforms the slot representations, and redistributes each token as a soft mixture over the transformed slots before adding the result back to the branch representations. This gives the two otherwise independent branches a shared bottleneck through which spatial and temporal evidence can be exchanged without top-k sparsity, suiting low-SNR EEG. Two companion mechanisms carry the rest of the argu

Load-bearing premise

The load-bearing assumption is that all ten foundation-model baselines are fairly evaluated through their official implementations and public checkpoints after being forced through STEAM's preprocessing (resampling to 250 Hz, interpolation to a 60-channel montage, and fixed segmentation windows); if that preprocessing disadvantages checkpoints not designed for it, the reported best-average-rank could reflect preprocessing mismatch rather than model quality.

What would settle it

Re-running the fourteen evaluation settings with each baseline using its native preprocessing, montage, and input length would decide the ranking claim: if STEAM's average rank against the ten baselines moves from 1.43 to, say, 3 or worse, the central claim fails. Separately, replacing the learned SSMoE routing with fixed random routing (same capacity) and observing no significant change in downstream accuracy would falsify the claim that shared soft routing is the mechanism driving the cross-branch gains.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • On the paper's reported evaluation, STEAM's Stage-I model is the best general EEG foundation model by average rank across these fourteen settings, at competitive inference cost.
  • Stage-II paradigm-specific continual pretraining consistently improves decoding over Stage-I alone on motor-imagery and emotion datasets, and reaches a given downstream accuracy 2.8x faster when initialized from Stage-I.
  • Lightweight adaptation that updates only the spatial and temporal embedding modules and the task head (about 5.1% of parameters) stays within roughly 1.5 points of full fine-tuning on most classification tasks, though not on the regression task.
  • Ablations show that removing either branch, the frequency gate, or the SSMoE module degrades performance, and that shared SSMoE beats a parameter-matched separate soft-MoE, indicating the shared routing space matters beyond capacity.
  • Removing either the reconstruction or the contrastive objective in Stage-I, or either objective in Stage-II, reduces downstream accuracy, so both self-supervised preservation and supervised specialization contribute.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: because SSMoE slots show differentiated temporal and channel routing, the same architecture could double as an interpretability probe—ablating individual slots or inspecting their routing weights could localize which channels and time intervals drive a prediction, a testable extension the paper does not perform.
  • Editorial inference: the hierarchical recipe is not EEG-specific; the same two-stage general-then-paradigm schedule could be tested on other physiological signals such as EMG or ECoG, where broad unlabeled corpora and paradigm-specific labeled datasets exist, to see whether the transfer-specialization trade-off also dissolves there.
  • Editorial inference: the appendix's rank comparison inherits the preprocessing pipeline (resampling to 250 Hz, interpolation to a 60-channel montage, fixed segmentation); a direct check would be to evaluate each baseline under its native montage and sampling rate to confirm that the 9.12-rank margin reflects model quality rather than preprocessing compatibility.
  • Editorial inference: the lightweight adaptation mode suggests deployment feasibility on resource-constrained devices, but the paper does not measure wall-clock or memory savings; a concrete next step would be to benchmark optimizer-state memory and per-trial latency on edge hardware.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes STEAM, a dual-branch spatio-temporal EEG encoder with a shared soft mixture-of-experts (SSMoE) cross-branch interaction module and frequency-aware gated attention. Pre-training is hierarchical: Stage-I self-supervised learning on TUH EEG, Stage-II paradigm-specific continual pre-training on MI or emotion corpora, followed by lightweight downstream adaptation that updates only ~5.1% of parameters. The central empirical claim is that Stage-I STEAM achieves the best average rank across seven downstream datasets and fourteen evaluation settings (Table 13) at competitive inference cost. The paper also reports consistent gains from Stage-II specialization and competitive lightweight adaptation. Extensive ablations cover components, pre-training objectives, fusion mechanisms, expert counts, and data ratios.

Significance. If the empirical claims are validated, STEAM would be a strong contribution to EEG foundation models: it addresses the general-vs-specialized trade-off with a concrete two-stage recipe, introduces a soft-MoE fusion mechanism with a compact slot bottleneck, and provides an unusually thorough set of ablations and hyperparameter details. The paper also ships architecture-level details (Tables 5–8) that support reproducibility. However, the headline 'best average rank' rests on an evaluation in which key architectural choices were selected using the same downstream benchmark, and the rank table excludes the very models that outperform STEAM on some settings. These issues must be resolved before the central claim can be accepted.

major comments (3)
  1. [Main Results; Fig. 6; Table 12; Table 13] The default SSMoE expert count (M=8) is chosen in Fig. 6 by measuring BAC on BNCI2015001, CHB-MIT, SEED, and BNCI2014009 under the few-shot protocols, and the Shared SSMoE fusion mechanism is selected using BNCI2015001 and SEED in Table 12. These exact datasets and protocols appear in the fourteen-setting aggregate-rank table (Table 13). Thus the architecture is tuned on the same benchmark used to compute the reported rank 1.43, making it an optimistic, in-sample estimate rather than an unbiased comparison. Please either freeze the architecture before touching the downstream benchmark, use a nested/held-out split for configuration selection, or demonstrate that the rank is robust across the full range of reasonable expert counts and fusion choices.
  2. [Appendix C, Table 13; Table 1] The aggregate rank excludes MIRepNet and STEAM-Spec, with the stated reason that neither reports all fourteen settings. However, Table 1 shows MIRepNet beats STEAM on BNCI2014001 in both protocols (LOSO 54.21 vs 53.43; few-shot 63.27 vs 59.66), and the abstract's claim of 'best average rank' is therefore relative only to a subset of models. BrainOmni is averaged over twelve settings because CHB-MIT results are unavailable, yet its rank 11.50 is reported alongside fourteen-setting averages as if comparable. The rank analysis should either include all models on the settings they support (with a principled missing-data rule) or state explicitly that the comparison is restricted to the ten FM baselines listed in Table 13; the current presentation overclaims the head-to-head result.
  3. [Appendix C, first paragraph; Experiment Settings] The paper states that 'all foundation-model baselines are evaluated using their official implementations and publicly released pre-trained checkpoints,' but all signals are resampled to 250 Hz, interpolated to a 60-channel montage, and segmented with fixed windows. No information is given about how each baseline's native tokenizer, positional encoding, or channel-order expectations were handled under this forced preprocessing. If a baseline was designed for a different montage or sampling rate, the preprocessing may systematically disadvantage it. Please report per-baseline preprocessing adaptations, or at least a sensitivity analysis on a subset, to rule out a preprocessing-mismatch explanation for the rank gaps.
minor comments (6)
  1. [Table 1] The row/column formatting is broken: '–38.02 ±0.03' and '60.83 ±0.30' appear in the BrainOmni and STEAM-Spec rows, respectively, with an ambiguous dash. These entries need to be aligned so that missing values ('–') are clearly distinguished from numeric results.
  2. [Fig. 4 and §Analysis of SSMoE Routing] The routing visualization is qualitative ('suggesting that they do not collapse'). Consider adding a quantitative diversity metric (e.g., average pairwise Jaccard overlap of top-routed tokens across slots) to support the non-collapse claim.
  3. [Eq. (14)] The negative-pair normalization over j includes i itself in the denominator; the standard InfoNCE formulation excludes the positive pair from the negative set. Please clarify whether the diagonal is masked; otherwise the loss is offset by a constant for each pair.
  4. [Table 10, few-shot footnote] The 'Few-shot (%)' column mixes very different fractions (e.g., 30% for BNCI2014001, ~12.75% for CHB-MIT, 60% for EEGMAT). The paper later uses the term 'few-shot' loosely; please state explicitly that these are per-dataset labeled-data fractions, not a uniform few-shot regime.
  5. [Appendix C, Aggregate-rank calculation] The rank formula R_m = |S_m|^{-1} sum_{s in S_m} r_{m,s} is given, but the number of settings per model differs because of missing CHB-MIT/BrainOmni. The table caption does not state how these missing entries affect the combined rank's variance; a footnote explaining the missing-data treatment would help.
  6. [Throughout] The paper uses 'Stage-I STEAM' and 'STEAM-Spec' inconsistently in the abstract and the main text; the abstract mentions 'STEAM' and 'hierarchical pre-training' without defining that STEAM-Spec is the Stage-II variant until the Experiments section. Please align terminology in the abstract.

Circularity Check

1 steps flagged

Benchmark-driven architecture selection partially undermines the headline rank; no definitional circularity in the pretraining derivation.

specific steps
  1. fitted input called prediction [Appendix C, 'SSMoE Expert-Count Sensitivity' (Fig. 6) and 'Controlled Comparison of Cross-Branch Fusion' (Table 12); Appendix C, 'Aggregate-rank calculation' (Table 13).]
    "Overall, the eight-expert configuration achieves strong and stable results across the evaluated datasets... We therefore adopted eight experts as the default SSMoE configuration. ... As shown in Table 12, Shared SSMoE achieved the best performance on both datasets"

    Fig. 6 selects the default SSMoE expert count by measuring downstream BAC on BNCI2015001, CHB-MIT, SEED, and BNCI2014009 under the few-shot protocols, and Table 12 selects the shared SSMoE fusion mechanism on BNCI2015001 and SEED. These exact dataset–scenario settings are then aggregated in Table 13 to produce STEAM's combined rank of 1.43. The headline 'best average rank' is therefore not an out-of-sample evaluation of a fixed architecture; the architecture's capacity and fusion design were chosen by inspecting the benchmarks being ranked. This is the fitted-input-called-prediction pattern: a configuration fitted to the evaluation data is presented as the benchmark's best performer. The result is partially circular/selection-dependent, though not equivalent by definition.

full rationale

The paper's central claim is empirical rather than a formal derivation. The two-stage pretraining objectives are well-defined, Stage-II corpora are disjoint from the downstream evaluation datasets (Appendix B), and the foundation-model baselines are evaluated through official implementations and public checkpoints. No definitional circularity or load-bearing self-citation chain is present: the self-citations (e.g., MIRepNet, the authors' benchmarking paper) are not used to justify STEAM's architecture, and excluding MIRepNet from the aggregate rank is transparently based on missing settings. However, the headline rank is compromised by benchmark-driven architecture selection: the default SSMoE expert count and the shared-fusion choice were selected using BAC on the same datasets that later appear in the aggregate-rank table. This does not make the architecture's output equivalent to its inputs by construction, but it means the reported 'best average rank' is partly a selection artifact rather than an unbiased evaluation. The paper's own ablations and external baselines give it substantial independent content, so the circularity is partial, not total.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 0 invented entities

The model is an empirical architecture whose central rank claim rests on data preprocessing, label alignment, and fair baseline evaluation, plus several hand-set hyperparameters. No new physical entities are introduced beyond standard architectural components.

free parameters (7)
  • Number of SSMoE expert slots (M) = 8
    Selected after measuring downstream BAC with M in {2,4,8,16} on four of the same benchmark datasets used for the headline rank claim (Fig. 6).
  • InfoNCE temperature (tau) = 0.20
    Fixed in both pre-training stages; no sensitivity analysis reported (Tables 7 and 8).
  • Masking ratio = 0.50
    Hand-set for masked reconstruction in Stage I and Stage II; no ablation across ratios.
  • Diversity loss weight (alpha) = 0.30
    Hand-set objective weight; no sensitivity analysis reported.
  • InfoNCE loss weight (lambda) = 0.01
    Hand-set objective weight; contrastive term is small relative to reconstruction and classification.
  • Classification loss weight (mu) = 1.00
    Hand-set for Stage-II supervised objective; no sensitivity analysis.
  • Embedding dimension (d) = 128
    Hand-chosen architectural capacity for both branches; no width or depth ablation reported.
axioms (4)
  • domain assumption Inverse-distance interpolation and zero-filling map arbitrary montages to the 60-channel model interface without systematic loss of discriminative spatial information
    Used for all Stage-I, Stage-II, and downstream datasets (Appendix B). If false, spatial tokens are partly synthetic and cross-dataset comparisons are distorted.
  • domain assumption The shared Stage-II label spaces are semantically consistent across datasets (feet/left/right motor imagery; negative/neutral/positive emotion)
    Stage-II supervised loss aggregates classification across datasets under shared labels (Eq. 17). Label mismatch would make specialization learn dataset artifacts.
  • domain assumption Pre-training on TUH clinical EEG transfers to MI, P300, seizure, cognition, emotion, and vigilance decoding
    Stage-I initialization is assumed to be a reusable backbone; the paper does not show which TUH properties drive transfer.
  • domain assumption Official checkpoints and implementations of baseline foundation models are compatible with STEAM's preprocessing and evaluation protocols
    All baselines are run by the authors with their official code (Appendix C); no raw outputs or configuration are released for verification.

pith-pipeline@v1.3.0-daily-deepseek · 21887 in / 14935 out tokens · 119708 ms · 2026-08-04T15:47:05.823467+00:00 · methodology

0 comments
read the original abstract

Brain-computer interfaces (BCIs) have been widely used in motor rehabilitation, disease diagnosis, and other neural engineering scenarios. However, conventional neural signal decoding algorithms often suffer from limited generalizability and high adaptation costs, motivating recent interest in BCI foundation models. Existing approaches still struggle to jointly achieve general transferability, accurate decoding, and efficient downstream adaptation. We present STEAM, a hierarchical transfer framework that reconciles general-purpose representation learning with paradigm-specific specialization in EEG foundation models. The framework is instantiated as a dual-branch spatio-temporal encoder in which a shared soft mixture-of-experts (SSMoE) module aligns the spatial and temporal branches, allowing complementary representations to exchange information through a compact set of soft slots. Across seven downstream datasets and fourteen evaluation settings, STEAM attains the best average rank among the compared methods at a competitive inference cost measured in FLOPs. Building upon the Stage-I general initialization, the hierarchical pre-training strategy further specializes the model to a target paradigm without retraining from scratch, yielding consistent gains in paradigm-specific decoding accuracy.

Figures

Figures reproduced from arXiv: 2608.02070 by Dingkun Liu, Dongrui Wu, Yuheng Chen, Zhu Chen.

Figure 1
Figure 1. Figure 1: Comparison of EEG foundation models by overall [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed STEAM framework. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Effect of Stage-I initialization on Stage-II paradigm [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: BAC differences relative to Pretrain-PEFT under [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Sensitivity of downstream BAC to the number of SSMoE experts under the dataset-specific few-shot protocols. [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Effect of the training data ratio on downstream BAC under full-model fine-tuning. [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Subject-level BAC on BNCI2014001 (n = 9). The upper and lower panels show LOSO and few-shot performance, respectively [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Subject-level BAC on BNCI2015001 (n = 12). The upper and lower panels show LOSO and few-shot performance, respectively [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Subject-level BAC on BNCI2014009 (n = 10). The upper and lower panels show LOSO and few-shot performance, respectively [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Subject-level BAC on CHB-MIT (n = 23). The upper and lower panels show LOSO and few-shot performance, respectively [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Subject-level BAC on SEED (n = 15). The upper and lower panels show LOSO and few-shot performance, respectively [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Subject-level BAC on EEGMAT (n = 36). The upper and lower panels show LOSO and few-shot performance, respectively [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Subject-level RMSE on SEED-VIG (n = 21). The upper and lower panels show LOSO and few-shot performance, respectively; lower values indicate better performance. Approach Scenario 14001 15001 14009 CHB-MIT SEED EEGMAT VIG Combined STEAM LOSO 1 1 1 1 1 1 1 1.43 Few-shot 1 1 2 4 1 1 3 CBraMod LOSO 2 5 5 11 2 8 17 6.50 Few-shot 6 8 10 2 3 3 9 Neuro-GPT LOSO 4 11 7 12 10 2 13 8.00 Few-shot 11 10 11 9 2 5 5 LaBr… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

56 extracted references · 3 linked inside Pith

  1. [1]

    Proceedings of the 9th international ACM SIGACCESS conference on Computers and accessibility , pages=

    Brain-computer interfaces (BCIs) for communication and control , author=. Proceedings of the 9th international ACM SIGACCESS conference on Computers and accessibility , pages=

  2. [2]

    The Journal of physiology , volume=

    Brain--computer interfaces: communication and restoration of movement in paralysis , author=. The Journal of physiology , volume=. 2007 , publisher=

    2007

  3. [3]

    Clinical neurophysiology , volume=

    Event-related EEG/MEG synchronization and desynchronization: basic principles , author=. Clinical neurophysiology , volume=. 1999 , publisher=

    1999

  4. [4]

    Journal of neural engineering , volume=

    A review of classification algorithms for EEG-based brain--computer interfaces , author=. Journal of neural engineering , volume=

  5. [5]

    IEEE transactions on rehabilitation engineering , volume=

    Optimal spatial filtering of single trial EEG during imagined hand movement , author=. IEEE transactions on rehabilitation engineering , volume=. 2000 , publisher=

    2000

  6. [6]

    IEEE Signal processing magazine , volume=

    Optimizing spatial filters for robust EEG single-trial analysis , author=. IEEE Signal processing magazine , volume=. 2008 , publisher=

    2008

  7. [7]

    2025 , publisher=

    Liu, Dingkun and Li, Siyang and Wang, Ziwei and Li, Wei and Wu, Dongrui , journal=. 2025 , publisher=

    2025

  8. [8]

    Human brain mapping , volume=

    Deep learning with convolutional neural networks for EEG decoding and visualization , author=. Human brain mapping , volume=. 2017 , publisher=

    2017

  9. [9]

    Journal of neural engineering , volume=

    EEGNet: a compact convolutional neural network for EEG-based brain--computer interfaces , author=. Journal of neural engineering , volume=. 2018 , publisher=

    2018

  10. [10]

    , journal =

    Craik, Alexander and He, Yongtian and Contreras-Vidal, Jose L. , journal =. Deep learning for electroencephalogram (. 2019 , publisher =

    2019

  11. [11]

    Advances in Neural Information Processing Systems , volume =

    Attention is all you need , author =. Advances in Neural Information Processing Systems , volume =

  12. [12]

    A multi-view

    Mane, Ravikiran and Robinson, Neethu and Vinod, A Prasad and Lee, Seong-Whan and Guan, Cuntai , booktitle=. A multi-view

  13. [13]

    IEEE Transactions on Neural Systems and Rehabilitation Engineering , volume=

    EEG conformer: Convolutional transformer for EEG decoding and visualization , author=. IEEE Transactions on Neural Systems and Rehabilitation Engineering , volume=. 2022 , publisher=

    2022

  14. [14]

    IEEE Journal of Biomedical and Health Informatics , volume=

    EEG-Deformer: A dense convolutional transformer for brain-computer interfaces , author=. IEEE Journal of Biomedical and Health Informatics , volume=. 2024 , publisher=

    2024

  15. [15]

    IEEE Journal of Biomedical and Health Informatics , year=

    DBConformer: Dual-branch convolutional transformer for EEG decoding , author=. IEEE Journal of Biomedical and Health Informatics , year=

  16. [16]

    Frontiers in Human Neuroscience , volume=

    BENDR: Using transformers and a contrastive self-supervised learning task to learn from massive amounts of EEG data , author=. Frontiers in Human Neuroscience , volume=. 2021 , publisher=

    2021

  17. [17]

    Advances in Neural Information Processing Systems , volume=

    Biot: Biosignal transformer for cross-data learning in the wild , author=. Advances in Neural Information Processing Systems , volume=

  18. [18]

    International Conference on Learning Representations , volume=

    Large brain model for learning generic representations with tremendous EEG data in BCI , author=. International Conference on Learning Representations , volume=

  19. [19]

    Advances in Neural Information Processing Systems , volume=

    Eegpt: Pretrained transformer for universal and reliable representation of eeg signals , author=. Advances in Neural Information Processing Systems , volume=

  20. [20]

    International conference on learning representations , volume=

    Cbramod: A criss-cross brain foundation model for eeg decoding , author=. International conference on learning representations , volume=

  21. [21]

    Advances in Neural Information Processing Systems , volume=

    Csbrain: A cross-scale spatiotemporal brain foundation model for eeg decoding , author=. Advances in Neural Information Processing Systems , volume=

  22. [22]

    Neural Networks , pages=

    Eegmamba: An eeg foundation model with mamba , author=. Neural Networks , pages=. 2025 , publisher=

    2025

  23. [23]

    International conference on learning representations , volume=

    NeuroLM: A universal multi-task foundation model for bridging the gap between language and EEG signals , author=. International conference on learning representations , volume=

  24. [24]

    Knowledge-Based Systems , pages=

    MIRepNet: A pipeline and pre-trained model for EEG-based motor imagery classification , author=. Knowledge-Based Systems , pages=. 2026 , publisher=

    2026

  25. [25]

    Deng, Guifeng and Wang, Pan and Wang, Jiquan and Rao, Shuying and Xie, Junyi and Guo, Wanjun and Li, Tao and Jiang, Haiteng , journal =

  26. [26]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Masked Autoencoders Are Scalable Vision Learners , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  27. [27]

    International Conference on Learning Representations , volume=

    From sparse to soft mixtures of experts , author=. International Conference on Learning Representations , volume=

  28. [28]

    Parameter-Efficient Transfer Learning for

    Houlsby, Neil and Giurgiu, Andrei and Jastrzebski, Stanislaw and Morrone, Bruna and De Laroussilhe, Quentin and Gesmundo, Andrea and Attariyan, Mona and Gelly, Sylvain , booktitle =. Parameter-Efficient Transfer Learning for. 2019 , organization =

    2019

  29. [29]

    Proceedings of the 2021 conference on empirical methods in natural language processing , pages=

    The power of scale for parameter-efficient prompt tuning , author=. Proceedings of the 2021 conference on empirical methods in natural language processing , pages=

    2021

  30. [30]

    Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina , booktitle =

  31. [31]

    Frontiers in neuroscience , volume=

    The temple university hospital EEG data corpus , author=. Frontiers in neuroscience , volume=. 2016 , publisher=

    2016

  32. [32]

    Biomedical Engineering/Biomedizinische Technik , volume=

    Random forests in non-invasive sensorimotor rhythm brain-computer interfaces: a practical and convenient non-linear classifier , author=. Biomedical Engineering/Biomedizinische Technik , volume=. 2016 , publisher=

    2016

  33. [33]

    GigaScience , volume=

    EEG datasets for motor imagery brain--computer interface , author=. GigaScience , volume=. 2017 , publisher=

    2017

  34. [34]

    Scientific Data , volume=

    A large EEG database with users’ profile information for motor imagery brain-computer interface research , author=. Scientific Data , volume=. 2023 , publisher=

    2023

  35. [35]

    GigaScience , volume=

    EEG dataset and OpenBMI toolbox for three BCI paradigms: An investigation into BCI illiteracy , author=. GigaScience , volume=. 2019 , publisher=

    2019

  36. [36]

    PloS one , volume=

    Evaluation of EEG oscillatory patterns and cognitive process during simple and compound limb motor imagery , author=. PloS one , volume=. 2014 , publisher=

    2014

  37. [37]

    PloS one , volume=

    A fully automated trial selection method for optimization of motor imagery based brain-computer interface , author=. PloS one , volume=. 2016 , publisher=

    2016

  38. [38]

    IEEE Transactions on biomedical engineering , volume=

    BCI2000: a general-purpose brain-computer interface (BCI) system , author=. IEEE Transactions on biomedical engineering , volume=. 2004 , publisher=

    2004

  39. [39]

    Frontiers in neuroscience , volume=

    Review of the BCI competition IV , author=. Frontiers in neuroscience , volume=. 2012 , publisher=

    2012

  40. [40]

    Journal of neural engineering , volume=

    Influence of P300 latency jitter on event related potential-based brain--computer interface performance , author=. Journal of neural engineering , volume=. 2014 , publisher=

    2014

  41. [41]

    2010 , month = jun, note =

    Guttag, John , title =. 2010 , month = jun, note =. doi:10.13026/C2K01R , url =

    doi:10.13026/c2k01r 2010

  42. [42]

    Data , volume=

    Electroencephalograms during mental arithmetic task performance , author=. Data , volume=. 2019 , publisher=

    2019

  43. [43]

    IEEE Transactions on autonomous mental development , volume=

    Investigating critical frequency bands and channels for EEG-based emotion recognition with deep neural networks , author=. IEEE Transactions on autonomous mental development , volume=. 2015 , publisher=

    2015

  44. [44]

    Journal of neural engineering , volume=

    A multimodal approach to estimating vigilance using EEG and forehead EOG , author=. Journal of neural engineering , volume=. 2017 , publisher=

    2017

  45. [45]

    Liu, Dingkun and Chen, Yuheng and Chen, Zhu and Cui, Zhenyao and Wen, Yaozhi and An, Jiayu and Luo, Jingwei and Wu, Dongrui , journal=

  46. [46]

    Thapa, Rahul and He, Bryan and Kj. Sleep. arXiv preprint arXiv:2405.17766 , year=

    Pith/arXiv arXiv

  47. [47]

    NeuroImage , volume=

    LMDA-Net: A lightweight multi-dimensional attention network for general EEG-based brain-computer interfaces and interpretability , author=. NeuroImage , volume=. 2023 , publisher=

    2023

  48. [48]

    2024 IEEE International Symposium on Biomedical Imaging (ISBI) , pages=

    Neuro-gpt: Towards a foundation model for eeg , author=. 2024 IEEE International Symposium on Biomedical Imaging (ISBI) , pages=. 2024 , organization=

    2024

  49. [49]

    arXiv preprint arXiv:2502.16060 , year=

    Tokenizing Single-Channel EEG with Time-Frequency Motif Learning , author=. arXiv preprint arXiv:2502.16060 , year=

    Pith/arXiv arXiv

  50. [50]

    Advances in Neural Information Processing Systems , volume=

    Brainomni: A brain foundation model for unified eeg and meg signals , author=. Advances in Neural Information Processing Systems , volume=

  51. [51]

    arXiv preprint arXiv:2509.17920 , year=

    SingLEM: Single-Channel Large EEG Model , author=. arXiv preprint arXiv:2509.17920 , year=

    Pith/arXiv arXiv

  52. [52]

    Advances in Neural Information Processing Systems , volume=

    LUNA: efficient and topology-agnostic foundation model for EEG signal analysis , author=. Advances in Neural Information Processing Systems , volume=

  53. [53]

    2022 44th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) , pages=

    Transformer convolutional neural networks for automated artifact detection in scalp EEG , author=. 2022 44th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) , pages=. 2022 , organization=

    2022

  54. [54]

    IEEE Transactions on Cognitive and Developmental Systems , year=

    Comparing Recognition Performance and Robustness of Multimodal Deep Learning Models for Multimodal Emotion Recognition , author=. IEEE Transactions on Cognitive and Developmental Systems , year=

  55. [55]

    IEEE transactions on affective computing , volume=

    Deap: A database for emotion analysis; using physiological signals , author=. IEEE transactions on affective computing , volume=. 2011 , publisher=

    2011

  56. [56]

    Autocalibration and Recurrent Adaptation: Towards a Plug and Play Online ERD-BCI , year=

    Faller, Josef and Vidaurre, Carmen and Solis-Escalante, Teodoro and Neuper, Christa and Scherer, Reinhold , journal=. Autocalibration and Recurrent Adaptation: Towards a Plug and Play Online ERD-BCI , year=