Pith. sign in

REVIEW 4 major objections 4 minor 23 references

The paper argues that flow-matching EEG generation is held back by a single global time schedule, and that letting each channel-time position advance at its own learned speed sharply improves fidelity and downstream gains.

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 01:29 UTC pith:ZUFCUJWE

load-bearing objection A genuinely new per-position time-warping mechanism for flow matching, with correct math and credible but partially self-consistent empirical evidence. the 4 major comments →

arxiv 2608.00048 v1 pith:ZUFCUJWE submitted 2026-07-25 eess.SP cs.AIcs.LG

Not All EEG Moments Are Equal: Position-Adaptive Time Scheduling for EEG Generation

Boheng Liu , Ziyu Li , Chenghua Duan , Qing Li , Xia Wu This is my paper
classification eess.SP cs.AIcs.LG
keywords EEG generationflow matchingposition-adaptive time schedulingdata augmentationbrain-computer interfacespectral consistency lossspatio-temporal attentionsynthetic EEG
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 claims that existing flow-matching EEG generators are held back because they advance every channel and time segment along the same global schedule, even though EEG contains rare, hard-to-reconstruct transients sitting on top of easier background rhythms. It proposes a Position-Adaptive Time Scheduling mechanism that learns, during training, a difficulty map from the model's own per-position reconstruction errors and uses it to give each channel-time position its own progress along the flow trajectory, while preserving the trajectory's endpoints. On three EEG datasets with very different recordings, this adaptive schedule and its two supporting components (factorized spatio-temporal attention and a frequency-aligned spectral consistency loss) reduce TS-FID by up to 62.2% and raise downstream classification accuracy gains by up to 6.77 percentage points over the strongest baseline. The central idea is that not all EEG moments are equal, and modeling that heterogeneity is the key to high-fidelity generation.

Core claim

The paper's central claim is that uniform time scheduling is the bottleneck in flow-matching EEG generation. The authors show that a position-specific time map τ_{i,j}(t) = t + m·s_{i,j}·t(1-t), built from a normalized exponential moving average s of per-position reconstruction error, preserves the flow endpoints (τ(0)=0, τ(1)=1) and monotonicity while pushing harder positions toward the clean distribution earlier. Coupled with a factorized spatio-temporal transformer that conditions on this map and a spectral loss reweighted by frequency to counter EEG's power-law bias, the framework reports the largest gains on the hardest channel-time positions, supporting the claim that heterogeneous dif

What carries the argument

Position-Adaptive Time Scheduling (PATS): an exponential moving average of per-channel, per-patch mean absolute reconstruction error is min-max normalized into a saliency map s, and a global time t becomes a matrix of position-specific times τ_{i,j}(t)=t + m·s_{i,j}·t(1-t). This map conditions a factorized spatio-temporal transformer via adaptive layer normalization, and enters the flow ODE through a corrected vector field that includes the position-dependent derivative (1 + m·s_{i,j}·(1-2t)). The identities τ(0)=0 and τ(1)=1 guarantee every position still interpolates from pure noise to the clean sample.

Load-bearing premise

The entire method rests on the assumption that the per-position reconstruction error averaged during training, computed with the model's own predictions, remains a stable and valid guide for which positions are hardest when generating new samples.

What would settle it

Train the full framework on one dataset, keep the network frozen, and at inference shuffle the learned saliency map s across positions (preserving its marginal distribution); if TS-FID and downstream accuracy do not degrade significantly, the position-specific schedule is not the active ingredient. Alternatively, train with a random, time-varying saliency map of the same magnitude; if it performs as well, the specific error-derived statistics are not needed.

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

If this is right

  • If the central claim holds, synthetic EEG produced with position-adaptive schedules should be more reliable for augmenting brain-computer interface training sets than uniform-schedule generation, across different montages and tasks.
  • The endpoint-preserving reparameterization is a drop-in modification for any flow matching model, so the same idea can be tested on other physiological signals with stereotyped transients, such as ECG or EMG.
  • The frequency-aligned spectral consistency loss implies that time-domain smoothness constraints systematically under-weight high-frequency EEG transients; spectral constraints with frequency reweighting should be preferred for power-law signals.
  • The paper's difficulty-stratified analysis predicts that gains concentrate on the hardest channel-time positions, which can be used as a diagnostic check in future generation studies.

Where Pith is reading between the lines

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

  • The learned difficulty map is a dataset-level statistic, so the method implicitly assumes that hard positions recur across subjects and sessions; a natural extension that the paper itself flags is instance-level difficulty that adapts to each sample's own transients.
  • Because the schedule is derived from the model's own errors, it forms a feedback loop; one testable extension is to recompute the difficulty map less frequently or with a separate teacher model to see whether the reported gains persist.
  • The TS-FID metric lives in a spectral feature space; a complementary check would measure preservation of clinically meaningful transient morphology, e.g., whether spike-like events survive generation better than with uniform schedules.

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

4 major / 4 minor

Summary. The paper proposes a flow-matching framework for EEG generation with Position-Adaptive Time Scheduling (PATS), Factorized Spatio-Temporal Attention, and a frequency-aligned multi-resolution spectral consistency loss. PATS maintains an EMA of per-channel/patch reconstruction errors, normalizes it to a saliency map, and uses it to define position-specific time progress τ_ij(t)=t+m s_ij t(1-t) for each channel-patch position while preserving the flow endpoints. The model is trained with a composite objective combining reconstruction, statistical, spectral, and correlation losses. Experiments on TUEV, BCIC-IV-2a, and SEED-IV compare against ESC-GAN, Diffusion-TS, PaD-TS, and JET, reporting large TS-FID reductions and downstream classification accuracy gains with both EEGNet and CBraMod. Leave-one-out ablations on TUEV show each component contributes, with PATS as the largest contributor.

Significance. If the attribution to the learned difficulty prior is correct, the paper makes a useful contribution by relaxing the uniform-time assumption in flow-based EEG generation and providing a concrete mechanism for allocating training signal to hard positions. The mathematical derivation of the position-specific vector field (Eq. 3) is sound, and the experimental package is substantial: three datasets with different protocols, two downstream classifiers, leave-one-out ablations, and a difficulty-stratified analysis. The downstream ΔAcc results, especially with EEGNet and CBraMod, are a genuine strength and provide external validation beyond the generation metric. However, the headline TS-FID fidelity claim is not independent of the training objective, and the causal role of the difficulty prior is not yet established by the current experiments.

major comments (4)
  1. [§4.3 and Eq. (6)] The primary fidelity metric, TS-FID, is computed in a spectral feature space (§4.3), while the training objective includes a multi-resolution STFT spectral consistency loss with frequency weighting (Eq. 6). The training loss and evaluation metric are therefore not independent; the reported TS-FID reductions, up to 62.2% on BCIC-IV-2a in Table 1, are at least partly a measure of how well the model minimizes a spectral objective aligned with the evaluator. This does not invalidate the downstream ΔAcc results, but it means the central 'fidelity' claim is inflated. Please report at least one independent fidelity metric (e.g., time-domain autocorrelation, power spectrum estimated with a different method, or FID from a pretrained encoder not used in training), and specify the exact TS-FID feature extractor to clarify its overlap with Eq. (6).
  2. [§3.2.1, Table 3, Fig. 4] The paper attributes the gains to the learned difficulty prior, but no control isolates the informative content of the saliency map s. A position-dependent schedule with a random or permuted s would also make τ non-uniform and could improve optimization; without such a control, the result does not establish that 'not all EEG moments are equal' is the operative cause. In addition, s is an EMA of the model's own errors during training, and the manuscript provides no analysis of the evolution or convergence of this EMA. The authors themselves note in the Conclusion and the Figure 4 discussion that s is a dataset-level statistic rather than an instance-level estimator. Please add (i) an ablation with shuffled or randomized s matched in marginal distribution, (ii) a plot of ē and s over training, and (iii) sensitivity to m, β, and γ. These are needed to support the central attribution claim.
  3. [§3.2.1, §4.3] The method depends on at least seven hyperparameters (m, β, γ, α, η, λ_stat, λ_spec, λ_corr, plus the number of STFT resolutions R), but the experimental section does not state their values or selection procedure. This is particularly important for PATS, since the central claim relies on m and β. Without these details, the results are not reproducible and the sensitivity of the method to the scheduler parameters is unknown. Please provide the full hyperparameter settings and, where feasible, a sensitivity analysis for the scheduler parameters.
  4. [§3.2.1 and §5] The use of training-time error statistics as a fixed inference-time schedule assumes the difficulty pattern of the generated distribution matches the training data. Because the model's errors change during training, the EMA may be stale or may reflect the model's current limitations rather than intrinsic data difficulty. The Conclusion's mention of future instance-level difficulty estimation is an implicit admission of this limitation, but it appears only there. Please quantify the distribution shift by comparing the train-time error map with reconstruction errors of the converged model on a validation set, and report whether the schedule remains calibrated at inference.
minor comments (4)
  1. [§3.1] The title and introduction emphasize 'moments,' but the schedule operates at the granularity of channel-patch positions; all P time points inside a patch share the same τ. With P=200 time points, this is a substantial simplification. Please state explicitly that the method is patch-level and discuss whether this granularity is adequate for the transient events that motivate the work.
  2. [§4.3] The TS-FID metric is described only vaguely as a Fréchet distance in a compact spectral feature space. Please define the exact feature extractor, the number of features, and the distance computation so readers can assess the metric and its relation to the spectral training loss.
  3. [Table 3] The ablation 'w/o Factorized Attn' does not state what architecture replaces the factorized spatio-temporal attention (e.g., full self-attention or a simpler conditioning mechanism). Please specify the replacement so the ablation is meaningful.
  4. [§4.5] All ablations are reported on TUEV only. While this is acceptable for a component analysis, a sentence acknowledging that the relative contributions may differ on the other two datasets would be appropriate.

Circularity Check

0 steps flagged

No significant circularity: PATS is an adaptive training loop and the main claims are grounded by external downstream classification.

full rationale

The core mechanism, Position-Adaptive Time Scheduling, maintains an EMA of per-position reconstruction errors and uses that statistic to define a position-specific flow-matching time map (Eq. 1) and loss weights (Eq. 4). This is a feedback loop, not a circular prediction: the saliency map is an emergent training statistic rather than a parameter fitted to the evaluation target, and the claim that hard positions improve is an empirical result rather than an identity. The paper's headline gains are also validated with external downstream classifiers (EEGNet in Table 1/Figure 3 and CBraMod in Table 2), so the central claim does not reduce to the training objective. The only partial overlap is that the spectral consistency loss (Eq. 6) operates in the spectral domain and TS-FID is a spectral-feature metric, with the loss band explicitly matched to the downstream evaluation protocol; this makes the fidelity metric partly aligned with training, but it does not collapse the scheduling contribution to the input, and the independent classification gains provide external grounding. No load-bearing self-citation or imported uniqueness theorem is present.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical entities. The main free parameters are the schedule margin, EMA decay, difficulty weighting, frequency reweighting exponent, and loss weights, none of which are derived from theory. The most important axiom is that the model's own EMA error is a stable and transferable difficulty statistic; this is the load-bearing premise of the proposed mechanism.

free parameters (7)
  • margin coefficient m = not stated
    Controls the maximum deviation from the uniform schedule in Eq. (1); chosen by hand/tuning; not fixed by theory.
  • EMA decay rate beta = not stated
    Determines how quickly the difficulty statistic updates; a free hyperparameter.
  • difficulty weight scaling gamma = not stated
    Scales the saliency-based weight in Eq. (4); ablation uses gamma=0.
  • frequency power alpha = not stated
    Exponent in w_k ∝ k^alpha for spectral reweighting; ablation uses alpha=0.
  • frequency band ratio eta = not stated
    Limits the frequency band over which reweighting is applied in Eq. (6).
  • loss weights lambda_stat, lambda_spec, lambda_corr = not stated
    Relative weights of the three auxiliary losses in Eq. (8).
  • number of STFT resolutions R = not stated
    Number of resolutions in the multi-resolution spectral loss.
axioms (4)
  • standard math Linear interpolation path x_t = t x_1 + (1-t) x_0 defines a valid flow-matching transport.
    Inherited from Lipman et al. 2023; used in Sec. 3.1 without re-derivation.
  • ad hoc to paper The EMA of per-position reconstruction error is a stable, dataset-level proxy for difficulty and remains valid at inference.
    Core design assumption in Sec. 3.2.1; no theoretical guarantee of stability or transfer to the generated distribution.
  • domain assumption Recurring transients and artifacts occupy consistent channel-time positions across the dataset.
    Motivates a dataset-level statistic rather than per-instance estimation; stated in Sec. 1 and Fig. 1.
  • domain assumption EEG spectral density follows a power law, motivating frequency reweighting with w_k ∝ k^alpha.
    Invoked in Sec. 3.3.3; no empirical verification on the three datasets is shown.

pith-pipeline@v1.3.0-alltime-deepseek · 11387 in / 8689 out tokens · 87720 ms · 2026-08-04T01:29:08.606521+00:00 · methodology

0 comments
read the original abstract

Electroencephalography (EEG) generation is essential for alleviating data scarcity and enabling large scale neural modeling in brain computer interface applications. However, existing flow based approaches assume that every channel and every time segment within a sample shares a single global time progression, overlooking the fact that not all EEG moments are equal. To address this overlooked heterogeneity, we propose an adaptive EEG generation framework built on conditional flow matching. The framework introduces Position-Adaptive Time Scheduling, which tracks per position reconstruction error to modulate a position specific time progress within the flow matching trajectory. It further incorporates Factorized Spatio-Temporal Attention and a frequency aligned multi resolution spectral consistency loss to model inter channel dependencies induced by volume conduction and compensate for the power law spectral bias of EEG, thereby improving the quality of generated signals. Extensive experiments on three EEG datasets with distinct acquisition protocols and task semantics show that our framework consistently outperforms the strongest baseline, reducing TS-FID by up to 62.2\% and improving downstream classification accuracy gain by up to 6.77 percentage points. These results suggest that the proposed method represents a promising step toward scalable, high fidelity data augmentation for real world brain computer interface applications.

Figures

Figures reproduced from arXiv: 2608.00048 by Boheng Liu, Chenghua Duan, Qing Li, Xia Wu, Ziyu Li.

Figure 1
Figure 1. Figure 1: Illustration of the overlooked heterogeneity in EEG generation. Quasi stationary background activity (green) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed framework. (a) Training pipeline: PATS derives a position specific time map from [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Downstream classification performance (Accuracy, Macro-F1, Macro-Recall) on (a) TUEV, (b) BCIC-IV-2a, [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Difficulty-stratified relative error reduction on TUEV. Instances are ranked by baseline (JET) error and [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] 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

23 extracted references · 3 canonical work pages

  1. [1]

    Müller-Putz, Alois Schlögl, and Gert Pfurtscheller

    Clemens Brunner, Robert Leeb, Gernot R. Müller-Putz, Alois Schlögl, and Gert Pfurtscheller. BCI competition 2008 – graz data set a. Dataset description, BCI Competition IV ,

    2008

  2. [6]

    URL https://openaccess.thecvf.com/content_CVPR_2019/html/ Karras_A_Style-Based_Generator_Architecture_for_Generative_Adversarial_Networks_CVPR_ 2019_paper.html

    doi:10.1109/CVPR.2019.00453. URL https://openaccess.thecvf.com/content_CVPR_2019/html/ Karras_A_Style-Based_Generator_Architecture_for_Generative_Adversarial_Networks_CVPR_ 2019_paper.html. 10 Position-Adaptive Time Scheduling for EEG GenerationA PREPRINT Vernon J. Lawhern, Amelia J. Solon, Nicholas R. Waytowich, Stephen M. Gordon, Chou P. Hung, and Brent...

    arXiv 2019

  3. [7]

    URL https://doi.org/10.1088/ 1741-2552/aace8c

    doi:10.1088/1741-2552/aace8c. URL https://doi.org/10.1088/ 1741-2552/aace8c. Tianhong Li and Kaiming He. Back to basics: Let denoising generative models denoise. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,

    doi:10.1088/1741-2552/aace8c

  4. [11]

    URL https://www.nature.com/articles/s41746-021-00440-5

    doi:10.1038/s41746-021-00440-5. URL https://www.nature.com/articles/s41746-021-00440-5. Kashif Rasul, Calvin Seward, Ingmar Schuster, and Roland V ollgraf. Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting. InProceedings of the 38th International Conference on Machine Learning, volume 139 ofProceedings of Mac...

    doi:10.1038/s41746-021-00440-5

  5. [12]

    URLhttps://doi.org/10.1109/TCDS.2024.3489357

    doi:10.1109/TCDS.2024.3489357. URLhttps://doi.org/10.1109/TCDS.2024.3489357. Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score- based generative modeling through stochastic differential equations. InInternational Conference on Learning Representations,

    arXiv 2024

  6. [14]

    Freek van Ede, Andrew J

    URL https://papers.neurips.cc/paper/2021/hash/ cfe8504bda37b575c70ee1a8276f3486-Abstract.html. Freek van Ede, Andrew J. Quinn, Mark W. Woolrich, and Anna C. Nobre. Neural oscillations: Sustained rhythms or transient burst-events?Trends in Neurosciences, 41(7):415–417,

    2021

  7. [15]

    URL https://doi.org/10.1016/j.tins.2018.04.004

    doi:10.1016/j.tins.2018.04.004. URL https://doi.org/10.1016/j.tins.2018.04.004. Julius Vetter, Jakob H. Macke, and Richard Gao. Generating realistic neurophysiological time series with denoising diffusion probabilistic models.Patterns, 5(9):101047,

    work page doi:10.1016/j.tins.2018.04.004 2018

  8. [18]

    URLhttps://www.nature.com/articles/s41597-025-04826-y

    doi:10.1038/s41597-025-04826-y. URLhttps://www.nature.com/articles/s41597-025-04826-y. Chaoqi Yang, M. Brandon Westover, and Jimeng Sun. BIOT: Biosignal transformer for cross- data learning in the wild. InAdvances in Neural Information Processing Systems, volume 36, pages 78240–78260,

    doi:10.1038/s41597-025-04826-y

  9. [19]

    Gunasekaran Yogarajan, Najah Alsubaie, G

    URL https://proceedings.neurips.cc/paper_files/paper/2023/hash/ f6b30f3e2dd9cb53bbf2024402d02295-Abstract-Conference.html. Gunasekaran Yogarajan, Najah Alsubaie, G. Rajasekaran, T. Revathi, Mohammed S. Alqahtani, Mohamed Abbas, Madshush M. Alshahrani, and Ben Othman Soufiene. EEG-based epileptic seizure detection using binary dragonfly algorithm and deep ...

    2023

  10. [20]

    URLhttps://www.nature.com/articles/s41598-023-44318-w

    doi:10.1038/s41598-023-44318- w. URLhttps://www.nature.com/articles/s41598-023-44318-w. Zhishui You, Yuzhu Guo, Xiulei Zhang, and Yifan Zhao. Virtual electroencephalogram acquisition: A review on electroencephalogram generative methods.Sensors, 25(10):3178,

    work page doi:10.1038/s41598-023-44318-

  11. [21]

    URL https: //www.mdpi.com/1424-8220/25/10/3178

    doi:10.3390/s25103178. URL https: //www.mdpi.com/1424-8220/25/10/3178. Xinyu Yuan and Yan Qiao. Diffusion-TS: Interpretable diffusion for general time series generation. InInternational Conference on Learning Representations, pages 41582–41610,

    work page doi:10.3390/s25103178

  12. [22]

    Zhi Zhang, Shenghua Zhong, and Yan Liu

    URL https://proceedings.iclr.cc/ paper_files/paper/2024/file/b5b66077d016c037576cc56a82f97f66-Paper-Conference.pdf. Zhi Zhang, Shenghua Zhong, and Yan Liu. Beyond mimicking under-represented emotions: Deep data augmentation with emotional subspace constraints for eeg-based emotion recognition. InProceedings of the AAAI Conference on Artificial Intelligenc...

    2024

  13. [23]

    URL https://ojs.aaai.org/index

    doi:10.1609/aaai.v38i9.28891. URL https://ojs.aaai.org/index. php/AAAI/article/view/28891. Wei-Long Zheng, Wei Liu, Yifei Lu, Bao-Liang Lu, and Andrzej Cichocki. EmotionMeter: A multimodal framework for recognizing human emotions.IEEE Transactions on Cybernetics, 49(3):1110–1122, 2019a. doi:10.1109/TCYB.2018.2797176. URLhttps://doi.org/10.1109/TCYB.2018.2...

    arXiv 2018

  14. [2012]

    URL https://www.frontiersin

    doi:10.3389/fnins.2012.00055. URL https://www.frontiersin. org/articles/10.3389/fnins.2012.00055/full. Yusuke Tashiro, Jiaming Song, Yang Song, and Stefano Ermon. CSDI: Conditional score-based diffusion models for probabilistic time series imputation. InAdvances in Neural Information Processing Systems, volume 34, pages 24804–24816. Curran Associates, Inc.,

    Pith/arXiv arXiv 2012

  15. [2016]

    URL https://www.frontiersin.org/articles/10.3389/fnins

    doi:10.3389/fnins.2016.00196. URL https://www.frontiersin.org/articles/10.3389/fnins. 2016.00196/full. William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195–4205,

    arXiv 2016

  16. [2017]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel

    URL https://papers.nips.cc/paper_files/paper/2017/hash/ 8a1d694707eb0fefe65871369074926d-Abstract.html. Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. InAdvances in Neural Information Processing Systems, volume 33, pages 6840–6851. Curran Associates, Inc.,

    2017

  17. [2018]

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter

    URLhttps://arxiv.org/abs/1806.01875. Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. InAdvances in Neural Information Processing Systems, volume

    Pith/arXiv arXiv

  18. [2019]

    URL https://www.frontiersin.org/articles/ 10.3389/fnhum.2019.00076/full

    doi:10.3389/fnhum.2019.00076. URL https://www.frontiersin.org/articles/ 10.3389/fnhum.2019.00076/full. Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. InAdvances in Neural Information Pro- cessing Systems, volume

    arXiv 2019

  19. [2021]

    URL https://doi.org/10

    doi:10.1109/TNNLS.2020.3016666. URL https://doi.org/10. 1109/TNNLS.2020.3016666. Meysam Golmohammadi, Amir Hossein Harati Nejad Torbati, Silvia Lopez de Diego, Iyad Obeid, and Joseph Picone. Automatic analysis of EEGs using big data and hybrid deep learning architectures.Frontiers in Human Neuroscience, 13:76,

    arXiv 2020

  20. [2023]

    URL https://openaccess.thecvf.com/content/ICCV2023/html/Peebles_Scalable_Diffusion_Models_ with_Transformers_ICCV_2023_paper.html

    doi:10.1109/ICCV51070.2023.00387. URL https://openaccess.thecvf.com/content/ICCV2023/html/Peebles_Scalable_Diffusion_Models_ with_Transformers_ICCV_2023_paper.html. Mathias Perslev, Sune Darkner, Lykke Kempfner, Miki Nikolic, Poul Jørgen Jennum, and Christian Igel. U-Sleep: Resilient high-frequency sleep staging.npj Digital Medicine, 4(1):72,

    arXiv 2023

  21. [2024]

    URL https: //www.cell.com/patterns/fulltext/S2666-3899(24)00189-2

    doi:10.1016/j.patter.2024.101047. URL https: //www.cell.com/patterns/fulltext/S2666-3899(24)00189-2. Jiquan Wang, Sha Zhao, Zhiling Luo, Yangxuan Zhou, Haiteng Jiang, Shijian Li, Tao Li, and Gang Pan. Cbramod: A criss-cross brain foundation model for eeg decoding. InInternational conference on learning representations, volume 2025, pages 75310–75346,

    arXiv 2024

  22. [2025]

    URLhttps://ojs.aaai.org/index.php/AAAI/article/view/34038

    doi:10.1609/aaai.v39i17.34038. URLhttps://ojs.aaai.org/index.php/AAAI/article/view/34038. Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. InInternational Conference on Learning Representations,

    doi:10.1609/aaai.v39i17.34038

  23. [2026]

    11 Position-Adaptive Time Scheduling for EEG GenerationA PREPRINT Banghua Yang, Fenqi Rong, Yunlong Xie, Du Li, Jiayang Zhang, Fu Li, Guangming Shi, and Xiaorong Gao

    URLhttps://arxiv.org/abs/ 2605.21280. 11 Position-Adaptive Time Scheduling for EEG GenerationA PREPRINT Banghua Yang, Fenqi Rong, Yunlong Xie, Du Li, Jiayang Zhang, Fu Li, Guangming Shi, and Xiaorong Gao. A multi-day and high-quality eeg dataset for motor imagery brain–computer interface.Scientific Data, 12(1):488,

    Pith/arXiv arXiv