Pith. sign in

REVIEW 4 major objections 7 minor 1 cited by

Geometric Machine Learning on EEG Signals

T0 review · 4 major / 7 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read EEG digit-thought decoding reaches 97% accuracy through a geometric manifold pipeline.

desk verdict The denoising numbers are plausible, but the headline 0.97 classification accuracy is not interpretable as a generalizable result because the non-digit class, the data split, and the hyperparameter selection protocol are never specified. read the letter →

arxiv 2502.05334 v2 pith:X3O224MV submitted 2025-02-07 cs.LG

classification cs.LG
keywords electroencephalographybrain-computerinterfacemanifoldlearningRicciflowgraphconvolutionalnetworkEEGdenoisingopen-setclassificationgeometricmachine
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

This paper tries to establish that high-dimensional, noisy EEG recordings contain a low-dimensional geometric structure tied to the content of a person's thoughts, and that this structure can be made explicit enough for classification. The proposed route is a two-stage pipeline: first an attention-based denoising network (AT-AT) removes muscle-artifact noise from individual channels, then a manifold-learning pipeline maps the denoised channels through a Fourier transform, Laplacian eigenmaps, a discrete Ricci-flow-like evolution, and a graph convolutional network before a lightweight classifier reads out the result. As a demonstration, the system classifies two-second samples of imagined digits versus non-digit thoughts with 97.0% test accuracy, and the denoising stage alone reaches a mean correlation above 0.95 with clean EEG at 2 dB SNR. A sympathetic reader would take the paper's contribution to be a concrete, parameter-light demonstration that geometric preprocessing can turn raw brainwave data into well-separated semantic classes.

What carries the argument

The load-bearing object is the discrete Ricci flow on a graph whose nodes are EEG channels. Starting from a graph built from one-sided FFT power spectra and Laplacian-eigenmap coordinates, each edge's Ollivier-Ricci curvature compares two neighbor distributions by Wasserstein distance, and edge weights are updated by $w_{uv}(t+1) = w_{uv}(t)e^{-\alpha\kappa(u,v)}$, which shrinks edges inside dense communities and leaves inter-community edges relatively heavy. After ten unsupervised iterations, the top fraction of edges is cut and the surviving adjacency is averaged into one matrix that trains a two-layer graph convolutional network. Upstream, the AT-AT denoising network, a convolutional autoencoder with selective transformer reconstruction and adversarial training, does the work of raising signal quality so that this geometry can appear.

What would settle it

Split the MindBigData recordings into a tuning set and a fresh lock box, choose the edge-cut ratio and transformer threshold using only the tuning set, then run the published pipeline on the lock box; if the fresh accuracy falls below the reported 93.66% confidence lower bound, the claimed 97% open-set accuracy is not reproducible.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that applying a discrete Ollivier-Ricci curvature flow to a graph of EEG channels, after denoising and spectral embedding, reveals community structure in the brain's electrical activity, and that structure survives dimensionality reduction well enough that digit- versus non-digit thoughts separate clearly in the final embedding. The paper reports 0.97 accuracy, 0.968 F1, 0.971 AUC, and strong class separation on a 100-sample held-out test set, with the actual classification carried out by a 5.2K-parameter 1D CNN. The claim is not merely that a classifier can distinguish the classes, but that the geometry itself is learnable from the data: the same Ricci-flow pipeline exposed separable structure that was invisible in the raw t-SNE plot and became visible after AT-AT denoising.

Load-bearing premise

The whole 97% figure rests on the assumption that the pipeline's tuning choices, especially the 0.6 edge-cut ratio and the 0.8 transformer threshold, were fixed without ever looking at the 100-sample test set; if those choices were guided by the test data, the accuracy is not an unbiased estimate of generalization.

Editorial extensions

If this is right

  • If the 97% figure holds, open-set EEG decoding, rejecting unknown thoughts while recognizing known mental states, is possible from two seconds of non-invasive recording without a heavy subject-specific classifier.
  • The denoising result suggests transformer-based artifact removal can be small and fast enough for live retraining, with total training time reported at 249 seconds including preprocessing.
  • The pipeline's output embedding is low-dimensional ($\mathbb{R}^{256}$), so downstream classification costs almost nothing, which matters for portable BCI hardware.
  • The observed pre- and post-denoising t-SNE separation is evidence that semantic class structure can be recovered geometrically rather than by brute-force deep learning.
  • Combining the pipeline's digit versus non-digit gate with an existing inter-digit classifier yields an estimated roughly 90% accuracy across all ten digits in an open-set setting with balanced classes.

Reading between the lines

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

  • The paper does not test whether the 0.6 edge-cut ratio transfers across tasks or subjects; if it must be retuned per dataset, the geometric structure it exposes may be less universal than the pipeline suggests.
  • A natural extension the paper leaves implicit is applying the same pipeline to imagined speech, motor imagery, or emotion tasks; those experiments would show whether the discovered structure is thought-specific or merely task-specific.
  • The open-set framing treats all non-digit thoughts as one unknown class, so a stricter test with unseen digit-like mental states would clarify whether the 97% accuracy reflects genuine rejection of unknowns.
  • If the geometric separability is real, the pipeline could serve as a general BCI front-end that maps raw multichannel recordings into a compact semantic embedding before any task-specific classifier is trained.
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

4 major / 7 minor

Summary. The paper proposes a two-stage pipeline for EEG-based thought classification: (1) an upstream denoising model called AT-AT, an autoencoder-targeted adversarial transformer that selectively invokes a transformer based on a correlation heuristic, and (2) a downstream geometric manifold learning pipeline that combines a one-sided FFT, Laplacian eigenmaps, an Ollivier-Ricci curvature-based discrete Ricci flow, graph cutting, and a graph convolutional network (GCN) to produce embeddings for a lightweight 1D-CNN classifier. The pipeline is applied to an 'open-set' imagined digit versus non-digit classification task using the MindBigData dataset. The paper reports a mean test correlation coefficient of 0.951 at 2 dB on EEGdenoiseNet denoising and a 97.0% test accuracy (95% CI: 93.66, 100.0) on the digit/non-digit task using a 100-sample test set. The authors frame the results as preliminary and call for further validation.

Significance. If the central claim were rigorously supported, the geometric manifold learning pipeline would be a notable contribution to EEG decoding: it combines several nontrivial geometric ideas (Ollivier-Ricci curvature, discrete Ricci flow, Laplacian eigenmaps) in a way that, the authors argue, reveals low-dimensional structure in high-dimensional EEG. The upstream AT-AT denoising system with a relatively small parameter count (438K+) is also of interest. However, the reported classification result rests on an evaluation protocol that is not fully specified and on hyperparameter choices that may have been made using the test data. The lack of a baseline comparator for the classification task, the undefined non-digit class, and the unexplained test-split procedure mean that the 97% accuracy cannot currently be interpreted as evidence of thought decoding. The paper is transparent about its preliminary nature and acknowledges the need for broader validation, which is a strength, but the methodological gaps are load-bearing rather than cosmetic.

major comments (4)
  1. [Section 2.2 / Appendix B] The non-digit class is never defined. The paper frames the task as 'open-set digit- versus non-digit thought classification' (Section 3), but it does not state how non-digit samples were obtained from MindBigData, which is a dataset of digit-imagery EEG trials. Without a precise description of the non-digit trials (e.g., rest periods, other imagery tasks, inter-trial intervals, or synthetic examples), the reported 0.970 accuracy may reflect trivial differences in recording conditions, subject state, or task instructions rather than a meaningful open-set discrimination. This is a fundamental validity issue for the paper's central claim.
  2. [Appendix B] The test set is described only as 'a lock box 100-sample test set comprising 20% of the initial classification data.' The paper does not state whether the split was random, chronological, per-session, per-subject, or stratified, nor does it report the number of subjects or sessions in the training and test sets. If test samples are temporally adjacent to training samples from the same session, temporal autocorrelation and subject-specific artifacts can inflate accuracy substantially. The absence of split details makes the 95% CI and the reported accuracy difficult to evaluate.
  3. [Section 2.2] Key hyperparameters appear to be selected from the data without a validation protocol. The 0.6 edge-cut ratio is described as 'selected in light of the weight distribution' (Section 2.2), and the 0.8 transformer invocation threshold in Section 2.1 is presented as a heuristic but with no mention of a held-out validation set. Since the Ricci flow, edge cutting, and GCN embedding are computed on the data before the test split is described, the text does not rule out that test samples influenced these choices. This is a potential leakage path that directly affects the 97% accuracy claim. A proper nested cross-validation or a clear statement that all hyperparameters were fixed on training data only is required.
  4. [Section 3 / Table 1] The classification result has no baseline comparison. The paper reports 97% accuracy for the full geometric pipeline but does not compare against, for example, the raw EEG signal, FFT alone, Laplacian eigenmaps alone, a standard GCN without Ricci flow, or a conventional CNN on raw EEG. Without such baselines, it is impossible to determine whether the geometric components (Ricci flow, edge cutting, Laplacian eigenmaps) contribute to the reported accuracy or whether a trivial classifier would achieve similar results on the same task. The denoising comparison in Table 1 also omits baselines for the -7 dB condition for GCTNet and EEGIFNet, and the reported metrics are not aligned across all rows, limiting the strength of the 'competitive performance' claim.
minor comments (7)
  1. [Section 2.1 / Appendix C] The 0.8 correlation threshold for transformer invocation is stated without a sensitivity analysis or a justification beyond a heuristic; a figure showing how performance varies with this threshold would improve reproducibility.
  2. [Appendix A, Algorithm 1] The custom distance metric in Algorithm 1 contains undefined notation: 'df(v)' appears to be a typo, and the formula mixes norms and FFT power spectra without specifying the exact dimensions or normalization. Please clarify.
  3. [Figure 2] The text says 'Figure 2, panel 2' when referring to the FFT/Laplacian eigenmap stage, but the figure appears to have six numbered panels; the reference is ambiguous.
  4. [References] The reference to Wu et al. (2024) for open-set classification is incorrect: that paper is about open-vocabulary learning, not open-set recognition. A more appropriate reference is needed for the open-set classification definition.
  5. [General] No code, data, or reproducibility statement is provided, despite the claim to 'ensure the integrity and reproducibility of our results' in Appendix B. Scripts or a detailed protocol would be necessary to evaluate the pipeline's behavior.
  6. [Section 2.2] The paper says '64,000 adjacency relationships and 1,000 hypothesized subgraphs' but does not explain how these numbers were chosen or how subgraphs are defined; this is a non-trivial design choice that should be justified.
  7. [Table 2] The confusion matrix is referred to in Table 2 but is not shown; please either include the matrix or remove the column header.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the headline results are benchmarked against external data and a held-out lock box; self-citations are contextual and not load-bearing.

full rationale

The paper's two headline results—denoising CC ≈ 0.95 on EEGdenoiseNet and 97% digit/non-digit classification on MindBigData—are empirical evaluations against external datasets, not derivations from the methods' own definitions. AT-AT is trained on EEGdenoiseNet before being applied to MindBigData, so the downstream classification is not forced by the denoising fit. The manifold pipeline (FFT, Laplacian eigenmap, Ollivier-Ricci flow, GCN) is described algorithmically, and the final 1D-CNN is evaluated on a 100-sample lock box (Appendix B). The 0.6 edge-cut ratio is 'selected in light of the weight distribution,' but the paper does not define the reported accuracy as a function of that ratio; this is a possible evaluation-protocol ambiguity and a correctness risk, not a circular reduction. The paper cites the authors' own prior work for AT-AT lineage, the LC SNR model, and adversarial-training details, but those are supporting descriptions rather than the sole justification of the headline numbers, which are also benchmarked against external baselines (Table 1). No equation in the paper defines a predicted quantity as identical to a fitted input, and no load-bearing claim rests solely on a self-citation chain. Hence no circular step can be exhibited under the required standard.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

No new physical entities are proposed. The central claim rests on a large number of hand-chosen hyperparameters (cut ratio, thresholds, iteration counts) and on domain assumptions about the datasets and the geometric hypothesis.

free parameters (8)
  • Edge cut ratio rho = 0.6
    Selected in Section 2.2 as 'optimal ... in light of the weight distribution', a post hoc choice that directly determines the adjacency matrix fed to the GCN.
  • Ricci flow learning rate alpha = not reported
    Controls the edge weight update w_uv(t+1)=w_uv(t)*exp(-alpha*kappa(u,v)) in Section 2.2; no value or sensitivity analysis is given.
  • Transformer invocation threshold = 0.8 (correlation)
    In Section 2.1, low-correlation sections (<0.8) are flagged for transformer reconstruction; this threshold is set without validation.
  • Number of Ricci flow iterations = 10
    Applied 'over 10 unsupervised iterations per subgraph' (Section 2.2); no convergence criterion or sensitivity analysis.
  • Number of hypothesized subgraphs = 1000
    Fixed in Section 2.2 as part of the graph initialization; no justification.
  • Initial adjacency relationships = 64000
    Fixed in Section 2.2; no justification.
  • Adversarial training cycles = 5
    Five cycles of adversarial training incorporated for spectral fidelity (Section 2.1); chosen by hand.
  • GCN embedding dimension = 256
    Final node embedding dimension in R256 for downstream classification (Appendix A.3); chosen without analysis.
assumptions (5)
  • domain assumption EEG data contain low-dimensional geometric structure that can be captured by graph-based manifold learning.
    Stated as 'an underlying hypothesis guiding this paper' in Section 1.
  • domain assumption The MindBigData dataset provides valid labels and the non-digit samples form a meaningful unknown class.
    The paper does not describe what the non-digit samples are or how they were selected (Section 2.1, Appendix B).
  • ad hoc to paper The custom distance metric in Algorithm 1 is a valid measure of channel relationships.
    Defined only in Appendix A.2 with undefined symbols (df(v)) and a typo; no empirical validation.
  • standard math Standard mathematical tools: Laplacian eigenmaps, FFT, Ollivier-Ricci curvature, GCNs.
    Used as building blocks; their correctness is assumed from prior literature.
  • domain assumption Semi-synthetic EEG-EMG mixtures from EEGdenoiseNet are representative of realistic EEG noise.
    AT-AT is trained and evaluated only on this benchmark (Section 2.1), so real-world performance is unknown.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Geometric Machine Learning on EEG Signals." pith.science (2026). https://pith.science/paper/X3O224MV

@misc{pith2026250205334,
  author       = {Pith},
  title        = {Pith review of: Geometric Machine Learning on EEG Signals},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X3O224MV}},
  note         = {Machine review of arXiv:2502.05334}
}
read the original abstract

Brain-computer interfaces (BCIs) offer transformative potential, but decoding neural signals presents significant challenges. The core premise of this paper is built around demonstrating methods to elucidate the underlying low-dimensional geometric structure present in high-dimensional brainwave data in order to assist in downstream BCI-related neural classification tasks. We demonstrate two pipelines related to electroencephalography (EEG) signal processing: (1) a preliminary pipeline removing noise from individual EEG channels, and (2) a downstream manifold learning pipeline uncovering geometric structure across networks of EEG channels. We conduct preliminary validation using two EEG datasets and situate our demonstration in the context of the BCI-relevant imagined digit decoding problem. Our preliminary pipeline uses an attention-based EEG filtration network to extract clean signal from individual EEG channels. Our primary pipeline uses a fast Fourier transform, a Laplacian eigenmap, a discrete analog of Ricci flow via Ollivier's notion of Ricci curvature, and a graph convolutional network to perform dimensionality reduction on high-dimensional multi-channel EEG data in order to enable regularizable downstream classification. Our system achieves competitive performance with existing signal processing and classification benchmarks; we demonstrate a mean test correlation coefficient of >0.95 at 2 dB on semi-synthetic neural denoising and a downstream EEG-based classification accuracy of 0.97 on distinguishing digit- versus non-digit- thoughts. Results are preliminary and our geometric machine learning pipeline should be validated by more extensive follow-up studies; generalizing these results to larger inter-subject sample sizes, different hardware systems, and broader use cases will be crucial.

Figures

Figures reproduced from arXiv: 2502.05334 by the authors.

Figure 1
Figure 1. AT-AT processing facilitates latent space separability of neural signal classes. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. The six manifold learning pipeline stages. (1) AT-AT filtration; (2) FFT upscaling [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Class embeddings produced by the GCN after the manifold learning pipeline. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Summary of Ollivier-Ricci curvature pipeline metrics as described in Section [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: Estimated true performance on the open-set imagined digit recognition problem [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Architectural diagram of the autoencoder-targeted adversarial transformer. The [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: AT-AT model workflow; also given in Choi (2025). From top to bottom: (1) raw input signal, (2) initial autoencoder filtration pass with high-noise target site masking, (3) time-series transformer-based reconstruction of target sites, (4) ground truth signal. connected …

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. Removing Neural Signal Artifacts with Autoencoder-Targeted Adversarial Transformers (AT-AT)

    cs.LG 2025-02 conditional novelty 5.0 of 10

    An autoencoder-gated adversarial transformer denoises EEG-EMG mixtures with reconstruction accuracy comparable to larger published models at a fraction of the model size.

Reference graph

Works this paper leans on

46 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    Anuragi, D

    A. Anuragi, D. S. Sisodia, and R. B. Pachori. Mitigating the curse of dimensionality using feature projection techniques on electroencephalography datasets: An empirical review. Artificial Intelligence Review, 57 0 (3): 0 75, February 2024

  2. [2]

    Armonaite, L

    K. Armonaite, L. Conti, L. Laura, M. Primavera, and F. Tecchio. Neural signal analysis with non-linear methods. Neural Signal Analysis, 2024

  3. [3]

    Belkin and P

    M. Belkin and P. Niyogi. Laplacian eigenmaps for dimensionality reduction and data representation. Neural Computation, 15 0 (6): 0 1373--1396, June 2003

  4. [4]

    X. Chen, X. Xu, A. Liu, S. Lee, X. Chen, X. Zhang, M. J. McKeown, and Z. J. Wang. Removal of muscle artifacts from the EEG : A review and recommendations. IEEE Sensors Journal, 19 0 (14): 0 5353--5368, March 2019

  5. [5]

    S. Cho, S. Cho, S. Park, H. Lee, H. Lee, and M. Lee. Curve your attention: Mixed-curvature transformers for graph representation learning. Technical report, arXiv preprint arXiv:2309.04082, 2023

  6. [6]

    B. J. Choi. Removing neural signal artifacts with autoencoder-targeted adversarial transformers ( AT-AT ). arXiv preprint, arXiv:2502.05332, 2025. doi:10.48550/arXiv.2502.05332. https://arxiv.org/abs/2502.05332

  7. [7]

    B. J. Choi and J. Liu. A low-cost transhumeral prosthesis operated via an ML -assisted EEG -head gesture control system. Journal of Neural Engineering, 22 0 (1), 2025. doi:10.1088/1741-2552/adae35

  8. [8]

    B. J. Choi, G. Milsap, C. A. Scholl, F. Tenore, and M. Ogg. Targeted adversarial denoising autoencoders ( TADA ) for neural time series filtration. arXiv preprint, arXiv:2501.04967, 2025. doi:10.48550/arXiv.2501.04967. https://arxiv.org/abs/2501.04967

Show all 46 references
  1. [9]

    Cinel, D

    C. Cinel, D. Valeriani, and R. Poli. Neurotechnologies for human cognitive augmentation: Current state of the art and future prospects. Frontiers in Human Neuroscience, 13, January 2019. doi:10.3389/fnhum.2019.00013

  2. [10]

    H. Cui, C. Li, A. Liu, R. Qian, and X. Chen. A dual-branch interactive fusion network to remove artifacts from single-channel EEG . IEEE Transactions on Instrumentation and Measurement, 73: 0 1--12, 2024. Art no. 4001912

  3. [11]

    Hemmelmann, M

    C. Hemmelmann, M. Horn, S. Reiterer, B. Schack, T. Süsse, and S. Weiss. Multivariate tests for the evaluation of high-dimensional EEG data. Journal of Neuroscience Methods, 139 0 (1): 0 111--120, October 2004

  4. [12]

    Huang, G

    J. Huang, G. Li, H. Su, and Z. Li. Development and continuous control of an intelligent upper-limb neuroprosthesis for reach and grasp motions using biological signals. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 52 0 (6): 0 3431--3441, June 2022. doi:10.1109/...

  5. [13]

    Huang, Y

    W. Huang, Y. Rong, T. Xu, F. Sun, and J. Huang. Tackling over-smoothing for general graph convolutional networks. Technical report, arXiv preprint arXiv:2008.09864, August 2020

  6. [14]

    N. M. Leite, E. T. Pereira, E. C. Gurjao, and L. R. Veloso. Deep convolutional autoencoder for EEG noise filtering. In 2018 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pages 2605--2612. IEEE, December 2018

  7. [15]

    J. León, J. J. Escobar, J. González, J. Ortega, F. M. Arrabal-Campos, A. Ortiz, and M. Damas. Recurrent neural networks and efficiency in high-dimensional EEG classification. In International Conference on Bioengineering and Biomedical Signal and Image Processing, pages 297--310, 2021

  8. [16]

    Liang, R

    Z. Liang, R. Zhou, L. Zhang, L. Li, G. Huang, Z. Zhang, and S. Ishii. EEGFuseNet : Hybrid unsupervised deep feature characterization and fusion for high-dimensional EEG with an application to emotion recognition. IEEE Transactions on Neural Systems and Rehabilitation Engineeri...

  9. [17]

    M. D. Linderman, G. Santhanam, C. T. Kemere, V. Gilja, S. O'Driscoll, M. Y. Byron, A. Afshar, S. I. Ryu, K. V. Shenoy, and T. H. Meng. Signal processing challenges for neural prostheses. IEEE Signal Process. Mag., 25: 0 18--28, 2008. doi:10.1109/MSP.2008.4408439

  10. [18]

    R. Liu, M. Xu, Y. Zhang, E. Peli, and A. D. Hwang. A pilot study on EEG -based evaluation of visually induced motion sickness. The Journal of Imaging Science and Technology, 64 0 (2): 0 20501--1, 2020

  11. [19]

    Lopez-Bernal, D

    D. Lopez-Bernal, D. Balderas, P. Ponce, and A. Molina. A state-of-the-art review of EEG -based imagined speech decoding. Front. Hum. Neurosci., 16, April 2022. doi:10.3389/fnhum.2022.867281

  12. [20]

    Mahapatra and P

    N. Mahapatra and P. Bhuyan. EEG -based classification of imagined digits using a recurrent neural network. Journal of Neural Engineering, 20 0 (2), 2023

  13. [21]

    Makeig, C

    S. Makeig, C. Kothe, T. Mullen, N. Bigdely-Shamlo, Z. Zhang, and K. Kreutz-Delgado. Evolving signal processing for brain–computer interfaces. Proc. IEEE, 100: 0 1567--1584, May 2012. doi:10.1109/JPROC.2012.2185009

  14. [22]

    Mishra, K

    R. Mishra, K. Sharma, and A. Bhavsar. Visual brain decoding for short duration EEG signals. In 2021 29th European Signal Processing Conference (EUSIPCO), pages 1226--1230. IEEE, August 2021

  15. [23]

    S. K. Mody and G. Rangarajan. Sparse representations of high dimensional neural data. Scientific Reports, 12 0 (1): 0 7295, 2022

  16. [24]

    J. P. C. Moreira, V. R. Carvalho, E. M. A. M. Mendes, A. Fallah, T. J. Sejnowski, C. Lainscsek, and L. Comstock. An open-access EEG dataset for speech decoding: Exploring the role of articulation and coarticulation. In bioRxiv, 2022

  17. [25]

    S. D. Muthukumaraswamy. High-frequency brain activity and muscle artifacts in MEG / EEG : A review and recommendations. Frontiers in Human Neuroscience, 7, 2013. doi:10.3389/fnhum.2013.00138

  18. [26]

    Narasimhan, H

    S. Narasimhan, H. J. Chiel, and S. Bhunia. Ultra-low-power and robust digital-signal-processing hardware for implantable neural interface microsystems. IEEE Trans. Biomed. Circuits Syst., 5: 0 169--178, April 2011. doi:10.1109/TBCAS.2010.2076281

  19. [27]

    Ollivier

    Y. Ollivier. Ricci curvature of metric spaces. Comptes Rendus Mathematique, 345 0 (11): 0 643--646, 2007

  20. [28]

    Y. T. Pan, J. L. Chou, and C. S. Wei. Matt: A manifold attention network for EEG decoding. In Advances in Neural Information Processing Systems, volume 35, pages 31116--31129, December 2022

  21. [29]

    J. Peng, R. Lei, and Z. Wei. Beyond over-smoothing: Uncovering the trainability challenges in deep graph neural networks. Technical report, arXiv preprint arXiv:2408.03669, August 2024

  22. [30]

    M. A. Pfeffer, S. S. Ling, and J. K. Wong. Exploring the frontier: Transformer-based models in EEG signal analysis for brain-computer interfaces. Computers in Biology and Medicine, page 108705, June 2024

  23. [31]

    X. Pu, P. Yi, K. Chen, Z. Ma, D. Zhao, and Y. Ren. EEGDnet : Fusing non-local and local self-similarity for EEG signal denoising with transformer. Computers in Biology and Medicine, 151: 0 106248, 2022

  24. [32]

    R. T. Schirrmeister, J. T. Springenberg, L. D. Fiederer, M. Glasstetter, K. Eggensperger, M. Tangermann, F. Hutter, W. Burgard, and T. Ball. Deep learning with convolutional neural networks for EEG decoding and visualization. Hum. Brain Mapp., 38 0 (11): 0 5391--5420, November...

  25. [33]

    J. Sia, E. Jonckheere, and P. Bogdan. Ollivier-Ricci curvature-based method to community detection in complex networks. Scientific Reports, 9 0 (1): 0 9800, 2019

  26. [34]

    M. Z. Soroush, P. Tahvilian, M. H. Nasirpour, K. Maghooli, K. Sadeghniiat-Haghighi, S. V. Harandi, Z. Abdollahi, A. Ghazizadeh, and N. J. Dabanloo. EEG artifact removal using sub-space decomposition, nonlinear dynamics, stationary wavelet transform, and machine learning algori...

  27. [35]

    F. T. Sun, S. A. Desai, T. K. Tcheng, and M. J. Morrell. Changes in the electrocorticogram after implantation of intracranial electrodes in humans: The implant effect. Clinical Neurophysiology, 129 0 (3): 0 676--686, March 2018

  28. [36]

    Y. Tian, Z. Lubberts, and M. Weber. Curvature-based clustering on graphs. Technical report, arXiv preprint arXiv:2307.10155, 2023

  29. [37]

    B. v. d. Berg, S. v. Donkelaar, and M. Alimardani. Inner speech classification using EEG signals: A deep learning approach. In 2021 IEEE 2nd International Conference on Human-Machine Systems (ICHMS), pages 1--4, 2021. doi:10.1109/ICHMS53169.2021.9582457

  30. [38]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, 2017

  31. [39]

    Vivancos and F

    D. Vivancos and F. Cuesta. Mindbigdata 2022: A large dataset of brain signals. Technical report, arXiv preprint arXiv:2212.14746, 2022

  32. [40]

    Y. Wang, C. M. Ting, X. Gao, and H. Ombao. Exploratory analysis of brain signals through low dimensional embedding. In 2019 9th International IEEE/EMBS Conference on Neural Engineering (NER), pages 997--1002, March 2019

  33. [41]

    Weber, E

    M. Weber, E. Saucan, and J. Jost. Characterizing complex networks with Forman-Ricci curvature and associated geometric flows. Journal of Complex Networks, 5 0 (4): 0 527--550, 2017

  34. [42]

    J. Wu, X. Li, S. Xu, H. Yuan, H. Ding, Y. Yang, X. Li, J. Zhang, Y. Tong, X. Jiang, and B. Ghanem. Towards open vocabulary learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, February 2024

  35. [43]

    J. Yin, A. Liu, C. Li, R. Qian, and X. Chen. A GAN guided parallel CNN and transformer network for EEG denoising. IEEE Journal of Biomedical and Health Informatics, May 2023

  36. [44]

    K.-S. Yoo. Motion prediction using brain waves based on artificial intelligence deep learning recurrent neural network. J. Exerc. Rehabil., 19 0 (4): 0 219–227, August 2023. doi:10.12965/jer.2346242.121

  37. [45]

    Zhang, C

    H. Zhang, C. Wei, M. Zhao, Q. Liu, and H. Wu. A novel convolutional neural network model to remove muscle artifacts from EEG . 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1265--1269, 2021 a

  38. [46]

    Zhang, M

    H. Zhang, M. Zhao, C. Wei, D. Mantini, Z. Li, and Q. Liu. EEGdenoiseNet : A benchmark dataset for deep learning solutions of EEG denoising. J. Neural Eng., 18: 0 056057, 2021 b

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.