Pith. sign in

REVIEW 4 major objections 5 minor 46 references

DMF2Mel: A Dynamic Multiscale Fusion Network for EEG-Driven Mel Spectrogram Reconstruction

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper proposes DMF2Mel, a four-module neural network that reconstructs mel spectrograms of imagined continuous speech from EEG, reporting a Pearson correlation of 0.074 on known subjects and 0.048 on unseen subjects on the SparrKULee…

desk verdict A serious architecture paper whose headline results are compromised by test-set hyperparameter selection and missing error bars. read the letter →

arxiv 2507.07526 v3 pith:ROSACUPU submitted 2025-07-10 cs.SD eess.AS

Cunhang Fan , Sheng Zhang , Jingjing Zhang , Enrui Liu , Xinhui Li , Gangming Zhao , Zhao Lv This is my paper
classification cs.SDeess.AS
keywords EEGspeechdecodingimaginedmelspectrogramreconstructioncontrastivelearningstatespacemodelKolmogorov-Arnoldnetworkmulti-scalefusionauditory
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a new architecture, DMF2Mel, reconstructs the mel spectrogram of imagined continuous speech from non-invasive EEG more accurately than existing models. On the SparrKULee dataset it reports a Pearson correlation of 0.074 for known subjects, a 48% improvement over the strongest baseline, and 0.048 for unseen subjects, a 35% improvement. The gain is attributed to four cooperating modules that separate speech-related signals from noise, fuse multi-scale features, and model long temporal dependencies with linear complexity. If the result holds, it would push non-invasive brain-computer interfaces closer to decoding continuous speech for people who cannot speak.

What carries the argument

The load-bearing machinery is the four-module architecture and a three-part loss. DC-FAM uses contrastive foreground/background feature aggregation with local convolution and global attention to suppress noise and preserve transient signals. HAMS-Net, based on U-Net, uses adaptive dual-attention feedback (ADAF) to let high-level semantics guide low-level detail generation. SplineMap attention replaces linear projections with learnable spline basis functions through AGKAN, combining global attention with spline-based local fitting. convMamba applies a bidirectional state space model (BiMamba) with linear complexity to capture long-range temporal dependencies. The total loss is $\mathcal{L}_{total} = \mathcal{L}_p + \lambda \mathcal{L}_1 + \beta \mathcal{L}_{InfoNCE}$, where the Pearson loss aligns temporal structure, the L1 term enforces sparsity, and the symmetric InfoNCE term maximizes cross-modal mutual information between EEG and spectrogram features.

What would settle it

Re-run DMF2Mel on the SparrKULee test set using a model selected on a separate validation split (stories or subjects not used in any tuning) and compare the Pearson correlation against SSM2Mel; if the gap shrinks or disappears, the reported improvement is an artifact of test-set tuning.

Watch

Extended reading notes

Core claim

The central claim is that DMF2Mel is a dynamic multiscale fusion network that achieves state-of-the-art mel spectrogram reconstruction from EEG during imagined continuous speech, outperforming VLAAI, HappyQuokka, and SSM2Mel on the SparrKULee dataset. The architecture combines a Dynamic Contrastive Feature Aggregation Module (DC-FAM) that contrastively separates speech-related foreground from interfering noise, a U-Net-based Hierarchical Attention-Guided Multi-Scale Network (HAMS-Net) for cross-scale semantic fusion, a SplineMap attention mechanism built on an Adaptive Gated Kolmogorov-Arnold Network (AGKAN) that balances local spline fitting with global context, and a bidirectional state space module (convMamba) for long-sequence temporal modeling. Reported results are a Pearson correlation of 0.074 on held-out stories, 0.048 on held-out subjects, and a composite score of 0.0653, each exceeding the compared baselines.

Load-bearing premise

The results rest on hyperparameters and module counts that the authors tuned by evaluating on the held-out test stories, with no separate validation split reported, so the headline 0.074 may be a best-of-search value rather than an unbiased estimate.

Editorial extensions

If this is right

  • If the reported gains hold on independent evaluation, non-invasive EEG speech decoding moves from word-level to minute-level continuous speech reconstruction.
  • The ablation study shows that removing DC-FAM costs 45.9% of performance, indicating that foreground/background separation is the most important module in this architecture.
  • The composite score, weighting known and unknown subjects as two-thirds and one-third, offers a reproducible benchmark protocol for future EEG-to-spectrogram systems.
  • convMamba's linear complexity means the approach scales to longer EEG segments without the quadratic cost of standard attention.
  • The combination of spline-based nonlinear fitting and bidirectional state space modeling suggests a general recipe for low-SNR neural decoding beyond speech.

Reading between the lines

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

  • The hyperparameters and module counts were selected by evaluating on the held-out test stories with no separate validation split reported, so the 0.074 figure may reflect favorable selection; an independent test set would be needed to confirm the improvement generalizes.
  • The same architecture could be adapted to other neural decoding tasks such as auditory attention detection or imagined speech classification, where noise suppression and long-range temporal modeling are also central.
  • The absolute Pearson correlations remain low (around 0.07), so the practical intelligibility of the reconstructed mel spectrograms is still unknown; a listening test or speech-quality metric would clarify the real-world value.
  • A natural testable extension would be to use the same four-module design with the envelope-decoding baselines to see whether the gain is specific to mel spectrograms or transfers to simpler targets.
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 / 5 minor

Summary. The paper proposes DMF2Mel, a multiscale fusion network for EEG-to-mel-spectrogram reconstruction, combining an embedding strength modulator, a contrastive feature aggregation module, a hierarchical attention U-Net, spline-map attention with a Kolmogorov-Arnold component, and a bidirectional Mamba-based temporal model. On the SparrKULee dataset, it reports Pearson correlations of 0.074 for held-out stories and 0.048 for held-out subjects, corresponding to composite score 0.0653 and claimed SOTA over VLAAI, HappyQuokka, and SSM2Mel. The paper includes module ablations and hyperparameter studies.

Significance. If the reported results are reliable, DMF2Mel would be a useful step for EEG-driven speech spectrogram reconstruction, and the proposed components (especially the contrastive foreground/background separation and spline-based attention) are architecturally interesting. The paper also provides a public code link, which supports reproducibility of the model itself. However, the significance is substantially weakened by the evaluation protocol: the headline numbers appear to be selected on the test set, and the comparisons against re-implemented baselines lack error bars and seeds. These issues must be resolved before the SOTA claim can be accepted.

major comments (4)
  1. [Sections 5.1–5.2 and Tables 3–4] The test-set hyperparameter selection is load-bearing. The paper defines a training set and two test splits but never introduces a validation split. Table 4 selects the loss weights lambda=0.5 and beta=0.1 by the resulting Pearson correlation, and Table 3 selects the module counts (2 convMambas, 4 DC-FAMs, 6 HAMS-Nets, 6 SplineMap Attentions) the same way. Because the same held-out stories are used both for model selection and for reporting the final 0.074, the headline value is the maximum over the searched configuration space rather than an unbiased estimate of generalization. The authors should either fix hyperparameters on a separate validation split and report a single final test evaluation, or use nested cross-validation and report the distribution of test scores.
  2. [Table 1 and Section 6.1] The baseline comparison is not backed by variance information. The text states that VLAAI, HappyQuokka, and SSM2Mel were re-implemented in this study, but provides no code, seeds, or confidence intervals. The reported margin over SSM2Mel on held-out stories is only 0.008 (0.074 vs. 0.066), which is likely within run-to-run or re-implementation variation. Please report mean and standard deviation over multiple training runs, and release the baseline implementations or provide sufficient detail for exact reproduction.
  3. [Abstract, Section 1, and Section 7] The paper repeatedly frames the task as "imagined speech" decoding, but the SparrKULee dataset contains EEG responses of subjects listening to stories (Section 5.1). No imagined-speech data are used in the experiments. This mismatch does not invalidate the mel-reconstruction measurements, but it means the stated application claim about imagined speech is untested and should be removed or clearly qualified.
  4. [Table 3 and Section 6.2] The convMamba ablation row reports "-" for 6 and 8 modules, with the text attributing the absence to overfitting, yet no criterion or evidence for overfitting is given. Since the module-count grid search is part of the test-set selection problem noted above, this row needs a clearer explanation (e.g., what value would indicate overfitting, and why those runs are excluded).
minor comments (5)
  1. [Abstract] There is a missing space in "results.Code" in the abstract.
  2. [Title page] "Corrresponding author" is a typo for "Corresponding author."
  3. [Section 6.1] The text refers to "Figure 2" and "Figure 3" when discussing the violin plots, but the actual figures are numbered Figure 3 and Figure 4. Please correct the cross-references.
  4. [Equation (15)] The L1 term is described as enforcing sparsity, but the mean absolute error between predicted and true spectrograms does not induce sparsity in the prediction; it is a reconstruction loss. Consider rephrasing.
  5. [Section 6.2] The ablation study reports percentage improvements relative to the full model but gives no statistical significance or error bars; please add at least mean and standard deviation over seeds.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline 0.074 is a best-of-search value because module counts and loss weights are selected on the held-out test set; no validation split is reported.

  1. fitted input called prediction [Section 5.1, Section 5.2, Section 6.2 (Tables 3 and 4)]
    "The training set includes 70 subjects, and the test set contains the same 70 subjects’ responses to audio stimuli not present in the training set, which is referred to as 'held-out stories.' ... Table 3 focuses on the impact of core module quantity configurations on model performance, taking the optimal combination of '2 convMambas, 4 DC-FAMs, 6 HAMS-Nets, and 6 SplineMap Attentions' as the benchmark ... Table 4 shows that when L1 constraint is applied, the reconstruction similarity peaks at 0.074 when the weight λ is set to 0.5."

    The paper defines held-out stories as the test set and never describes a validation split. Tables 3 and 4 then choose module counts and loss weights by the resulting Pearson correlation, with the optimum equal to the headline 0.074. The reported 0.074 is therefore the maximum of a configuration search evaluated on the same test stories used for the final claim, not an unbiased held-out measurement. The comparison to re-implemented baselines inherits the same selection: the margin over SSM2Mel (0.008) is a best-of-search difference, not a validated effect.

full rationale

The architecture derivation itself is not definitionally circular: the module equations are standard and the loss function Lp = 1 - Pearson r is a legitimate training objective, even though it equals the evaluation metric. The circularity lies in the evaluation protocol. Section 5.1 labels held-out stories and held-out subjects as the test set, and Section 5.2 gives training details without any validation split. Section 6.2 then tunes module counts and loss weights (lambda and beta) directly by the Pearson correlation observed on those test configurations, selecting the combination that yields 0.074. Thus the central SOTA claim reduces to a fitted maximum over the searched hyperparameter grid, and the 35-48% improvement claims are not independent measurements. No self-citation chain or imported uniqueness theorem is load-bearing, so the score is 6 rather than 8-10.

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

The central claim depends on the evaluation protocol and the dataset. The paper does not report a validation split, so several free parameters are tuned on the test set. The remaining assumptions concern the benchmark, the baseline re-implementations, and the metric.

free parameters (4)
  • L1 loss weight lambda = 0.5 (best of 0, 0.1, 0.3, 0.5, 1)
    Tuned on held-out stories (Table 4); no validation split described.
  • InfoNCE loss weight beta = 0.1 (best of 0, 0.1, 0.3, 0.5, 1)
    Tuned on held-out stories (Table 4); no validation split described.
  • Module counts (convMamba, DC-FAM, HAMS-Net, SplineMap) = 2, 4, 6, 6
    Chosen as the best combination in Table 3 by scanning counts and evaluating on the test metric.
  • Training schedule (epochs, learning rate, scheduler, segment length) = 1000 epochs, lr 0.0005, StepLR 0.9/50 epochs, 5-second segments
    Standard training choices; sensitivity to these settings is not reported.
assumptions (4)
  • domain assumption SparrKULee is a valid benchmark for EEG-driven speech reconstruction.
    Used as the sole dataset without external validation of the ground truth.
  • domain assumption Re-implemented baselines faithfully represent VLAAI, HappyQuokka, and SSM2Mel.
    Table 1 footnote says these results are re-implemented, but no code or detailed protocol is provided.
  • domain assumption Pearson correlation of mel spectrograms is a meaningful metric for reconstruction quality.
    Used both as the evaluation metric and as a training loss.
  • domain assumption Listening EEG data can be treated as imagined speech decoding.
    The paper frames the task as imagined speech, but the dataset contains auditory responses during listening.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DMF2Mel: A Dynamic Multiscale Fusion Network for EEG-Driven Mel Spectrogram Reconstruction." pith.science (2026). https://pith.science/paper/ROSACUPU

@misc{pith2026250707526,
  author       = {Pith},
  title        = {Pith review of: DMF2Mel: A Dynamic Multiscale Fusion Network for EEG-Driven Mel Spectrogram Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ROSACUPU}},
  note         = {Machine review of arXiv:2507.07526}
}
read the original abstract

Decoding speech from brain signals is a challenging research problem. Although existing technologies have made progress in reconstructing the mel spectrograms of auditory stimuli at the word or letter level, there remain core challenges in the precise reconstruction of minute-level continuous imagined speech: traditional models struggle to balance the efficiency of temporal dependency modeling and information retention in long-sequence decoding. To address this issue, this paper proposes the Dynamic Multiscale Fusion Network (DMF2Mel), which consists of four core components: the Dynamic Contrastive Feature Aggregation Module (DC-FAM), the Hierarchical Attention-Guided Multi-Scale Network (HAMS-Net), the SplineMap attention mechanism, and the bidirectional state space module (convMamba). Specifically, the DC-FAM separates speech-related "foreground features" from noisy "background features" through local convolution and global attention mechanisms, effectively suppressing interference and enhancing the representation of transient signals. HAMS-Net, based on the U-Net framework,achieves cross-scale fusion of high-level semantics and low-level details. The SplineMap attention mechanism integrates the Adaptive Gated Kolmogorov-Arnold Network (AGKAN) to combine global context modeling with spline-based local fitting. The convMamba captures long-range temporal dependencies with linear complexity and enhances nonlinear dynamic modeling capabilities. Results on the SparrKULee dataset show that DMF2Mel achieves a Pearson correlation coefficient of 0.074 in mel spectrogram reconstruction for known subjects (a 48% improvement over the baseline) and 0.048 for unknown subjects (a 35% improvement over the baseline).Code is available at: https://github.com/fchest/DMF2Mel.

Figures

Figures reproduced from arXiv: 2507.07526 by the authors.

Figure 1
Figure 1. The overall framework of the DMF2Mel model, which consists of Embedding Strength Modulator (ESM), Dynamic Con [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The components of the ADAF block in HAMS-Net. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Violin plot of the mel spectrogram test results for [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 40 canonical work pages

  1. [1]

    Bernd Accou, Jonas Vanthornhout, Hugo Van hamme, and Tom Francart. 2023. Decoding of the speech envelope from EEG using the VLAAI deep neural network. Scientific Reports 13, 1 (2023), 812

    work page 2023

  2. [2]

    Hassan Akbari, Bahar Khalighinejad, Jose L Herrero, Ashesh D Mehta, and Nima Mesgarani. 2019. Towards reconstructing intelligible speech from the human auditory cortex. Scientific reports 9, 1 (2019), 874

    work page 2019

  3. [3]

    Gopala K Anumanchipalli, Josh Chartier, and Edward F Chang. 2019. Speech synthesis from neural decoding of spoken sentences. Nature 568, 7753 (2019), 493–498

    work page 2019

  4. [4]

    Lies Bollens, Bernd Accou, Marlies Gillis, Wendy Verheijen, Tom Francart, et al

    work page

  5. [5]

    Xupeng Chen, Ran Wang, Amirhossein Khalilian-Gourtani, Leyao Yu, Patricia Dugan, Daniel Friedman, Werner Doyle, Orrin Devinsky, Yao Wang, and Adeen Flinker. 2024. A neural speech decoding framework leveraging deep learning and speech synthesis. Nature Machine Intelligence 6, 4 (2024), 467–480

    work page 2024

  6. [6]

    Sahil Datta and Nikolaos V Boulgouris. 2021. Recognition of grammatical class of imagined words from EEG signals using convolutional neural network. Neu- rocomputing 465 (2021), 301–309

    work page 2021

  7. [7]

    Cunhang Fan, Ying Chen, Jian Zhou, Zexu Pan, Jingjing Zhang, Youdian Gao, Xiaoke Yang, Zhengqi Wen, and Zhao Lv. 2025. M3ANet: Multi-scale and Multi- Modal Alignment Network for Brain-Assisted Target Speaker Extraction. arXiv preprint arXiv:2506.00466 (2025)

    work page Pith review arXiv 2025

  8. [8]

    Cunhang Fan, Wang Xiang, Jianhua Tao, Jiangyan Yi, and Zhao Lv. 2025. Cross- Modal Knowledge Distillation With Multi-Stage Adaptive Feature Fusion for Speech Separation. IEEE Transactions on Audio, Speech and Language Processing (2025)

    work page 2025

Show all 46 references
  1. [9]

    Cunhang Fan, Xiaoke Yang, Hongyu Zhang, Ying Chen, Lu Li, Jian Zhou, and Zhao Lv. 2025. ListenNet: A Lightweight Spatio-Temporal Enhancement Nested Network for Auditory Attention Detection. arXiv preprint arXiv:2505.10348 (2025)

    2025 arXiv

  2. [10]

    Cunhang Fan, Hongyu Zhang, Wei Huang, Jun Xue, Jianhua Tao, Jiangyan Yi, Zhao Lv, and Xiaopei Wu. 2024. DGSD: Dynamical graph self-distillation for EEG-based auditory spatial attention detection. Neural Networks 179 (2024), 106580

    2024

  3. [11]

    Cunhang Fan, Hongyu Zhang, Qinke Ni, Jingjing Zhang, Jianhua Tao, Jian Zhou, Jiangyan Yi, Zhao Lv, and Xiaopei Wu. 2025. Seeing helps hearing: A multi-modal dataset and a mamba-based dual branch parallel network for auditory attention decoding. Information Fusion (2025), 102946

    2025

  4. [12]

    Cunhang Fan, Jingjing Zhang, Hongyu Zhang, Wang Xiang, Jianhua Tao, Xinhui Li, Jiangyan Yi, Dianbo Sui, and Zhao Lv. 2024. MSFNet: Multi-scale fusion network for brain-controlled speaker extraction. In Proceedings of the 32nd ACM International Conference on Multimedia . 1652–1661

    2024

  5. [13]

    Cunhang Fan, Sheng Zhang, Jingjing Zhang, Zexu Pan, and Zhao Lv. 2025. SSM2Mel: State Space Model to Reconstruct Mel Spectrogram from the EEG. arXiv preprint arXiv:2501.10402 (2025)

    2025 arXiv

  6. [14]

    Adam Gazzaley, Jeffrey W Cooney, Kevin McEvoy, Robert T Knight, and Mark D’esposito. 2005. Top-down enhancement and suppression of the magnitude and speed of neural activity. Journal of cognitive neuroscience 17, 3 (2005), 507–517

    2005

  7. [15]

    Bernhard Graimann, Brendan Allison, and Gert Pfurtscheller. 2010. Brain– computer interfaces: A gentle introduction. In Brain-computer interfaces: Revolu- tionizing human-computer interaction. Springer, 1–27

    2010

  8. [16]

    Shuai Hu, Feng Gao, Xiaowei Zhou, Junyu Dong, and Qian Du. 2024. Hybrid convolutional and attention network for hyperspectral image denoising. IEEE Geoscience and Remote Sensing Letters (2024)

    2024

  9. [17]

    Gautam Krishna, Co Tran, Yan Han, Mason Carnahan, and Ahmed H Tewfik. 2020. Speech synthesis using EEG. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 1235–1238

    2020

  10. [18]

    SangJong Lee, Jin-Kwang Kim, JunHo Kim, TaeHan Kim, and James Lee

  11. [19]

    Young-Eun Lee, Seo-Hyun Lee, Sang-Ho Kim, and Seong-Whan Lee. 2023. To- wards voice reconstruction from EEG during imagined speech. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 37. 6030–6038

    2023

  12. [20]

    Mengqi Lei, Haochen Wu, Xinhua Lv, and Xin Wang. 2024. ConDSeg: A General Medical Image Segmentation Framework via Contrast-Driven Feature Enhance- ment. arXiv preprint arXiv:2412.08345 (2024)

    2024 arXiv

  13. [21]

    Hao Li, Yuan Fang, Xueliang Zhang, Fei Chen, and Guanglai Gao. 2024. Cross- Attention-Guided WaveNet for EEG-to-MEL Spectrogram Reconstruction. In Proc. Interspeech 2024. 2620–2624

    2024

  14. [22]

    Yuanning Li, Claire Tang, Junfeng Lu, Jinsong Wu, and Edward F Chang. 2021. Human cortical encoding of pitch in tonal and non-tonal languages. Nature communications 12, 1 (2021), 1161

    2021

  15. [23]

    Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. 2024. Vmamba: Visual state space model. Advances in neural information processing systems 37 (2024), 103031–103063

    2024

  16. [24]

    Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljačić, Thomas Y Hou, and Max Tegmark. 2024. Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756 (2024)

    2024 arXiv

  17. [25]

    Nrushingh Charan Mahapatra and Prachet Bhuyan. 2023. EEG-based classifi- cation of imagined digits using a recurrent neural network. Journal of neural engineering 20, 2 (2023), 026040

    2023

  18. [26]

    Sean L Metzger, Jessie R Liu, David A Moses, Maximilian E Dougherty, Margaret P Seaton, Kaylo T Littlejohn, Josh Chartier, Gopala K Anumanchipalli, Adelyn Tu- Chan, Karunesh Ganguly, et al . 2022. Generalizable spelling using a speech neuroprosthesis in an individual with seve...

    2022

  19. [27]

    David A Moses, Nima Mesgarani, Matthew K Leonard, and Edward F Chang. 2016. Neural speech recognition: continuous phoneme decoding using spatiotemporal representations of human cortical activity. Journal of neural engineering 13, 5 (2016), 056004

    2016

  20. [28]

    David A Moses, Sean L Metzger, Jessie R Liu, Gopala K Anumanchipalli, Joseph G Makin, Pengfei F Sun, Josh Chartier, Maximilian E Dougherty, Patricia M Liu, Gary M Abrams, et al. 2021. Neuroprosthesis for decoding speech in a paralyzed person with anarthria. New England Journal...

    2021

  21. [29]

    Zhenyu Piao, Miseul Kim, Hyungchan Yoon, and Hong-Goo Kang. 2023. Hap- pyquokka system for icassp 2023 auditory eeg challenge. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 1–2

    2023

  22. [30]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolu- tional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceeding...

    2015

  23. [31]

    Ariel Tankus, Itzhak Fried, and Shy Shoham. 2012. Structured neuronal encoding and decoding of human speech features. Nature communications 3, 1 (2012), 1015

    2012

  24. [32]

    Bob Van Dyck, Liuyin Yang, and Marc M Van Hulle. 2023. Decoding auditory eeg responses using an adapted wavenet. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 1–2

    2023

  25. [33]

    Sarah K Wandelt, David A Bjånes, Kelsie Pejsa, Brian Lee, Charles Liu, and Richard A Andersen. 2024. Representation of internal speech by single neurons in human supramarginal gyrus. Nature human behaviour 8, 6 (2024), 1136–1149

    2024

  26. [34]

    Zhenhailong Wang and Heng Ji. 2022. Open vocabulary electroencephalography- to-text decoding and zero-shot sentiment classification. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 36. 5350–5358

    2022

  27. [35]

    Francis R Willett, Donald T Avansino, Leigh R Hochberg, Jaimie M Henderson, and Krishna V Shenoy. 2021. High-performance brain-to-text communication via handwriting. Nature 593, 7858 (2021), 249–254

    2021

  28. [36]

    Francis R Willett, Erin M Kunz, Chaofei Fan, Donald T Avansino, Guy H Wilson, Eun Young Choi, Foram Kamdar, Matthew F Glasser, Leigh R Hochberg, Shaul Druckmann, et al. 2023. A high-performance speech neuroprosthesis. Nature 620, 7976 (2023), 1031–1036

    2023

  29. [37]

    Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. 2018. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV) . 3–19

    2018

  30. [38]

    Kunpeng Xu, Lifei Chen, and Shengrui Wang. 2024. KAN4Drift: Are KAN Ef- fective for Identifying and Tracking Concept Drift in Time Series?. In NeurIPS Workshop on Time Series in the Age of Large Models

    2024

  31. [39]

    Xiran Xu, Bo Wang, Yujie Yan, Haolin Zhu, Zechen Zhang, Xihong Wu, and Jing Chen. 2024. ConvConcatNet: a deep convolutional neural network to reconstruct mel spectrogram from the EEG. In2024 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (I...

    2024

  32. [40]

    Sheng Yan, Cunhang Fan, Hongyu Zhang, Xiaoke Yang, Jianhua Tao, and Zhao Lv

  33. [41]

    Qi Yang, Jiajun Ji, Ruomiao Jing, Haifeng Su, Shuohong Wang, and Aike Guo

  34. [42]

    Quan Zhou, Changhua Pei, Fei Sun, Jing Han, Zhengwei Gao, Dan Pei, Haiming Zhang, Gaogang Xie, and Jianhui Li. 2024. KAN-AD: Time series anomaly detection with Kolmogorov-Arnold networks. arXiv preprint arXiv:2411.00278 (2024)

    2024 arXiv

  35. [43]

    Advances in Neural Information Processing Systems 37 (2024), 31688–31707

    Darnet: Dual attention refinement network with spatiotemporal construc- tion for auditory attention detection. Advances in Neural Information Processing Systems 37 (2024), 31688–31707

    2024

  36. [2023]

    SparrKULee: A speech-evoked auditory response repository of the KU Leuven, containing EEG of 85 participants. (2023)

    2023

  37. [2024]

    arXiv preprint arXiv:2410.14939 (2024)

    Hippo-kan: Efficient kan model for time series analysis. arXiv preprint arXiv:2410.14939 (2024)

    2024 arXiv

  38. [2025]

    Scientific Reports 15, 1 (2025), 6982

    Reynolds rules in swarm fly behavior based on KAN transformer tracking method. Scientific Reports 15, 1 (2025), 6982

    2025

Pith tools

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