Pith. sign in

REVIEW 4 major objections 6 minor 19 references

The paper claims that earphone-recorded ExG signals, decomposed into 12 physiology-informed frequency sub-bands and pre-trained with reconstruction objectives on unlabeled free-living data, yield a single task-agnostic representation that b

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 ↗

A 12-band physiology-informed tokenizer with a bidirectional Mamba encoder, pre-trained on free-living earphone ExG data, beats task-specific baselines across five-sense and public ExG benchmarks.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Real hardware-plus-data contribution, but the 'five senses' generalization claim rests on an evaluation split that leaks subject/session identity and a protocol that co-varies with facial EMG; the public-benchmark evidence is stronger and worth referee time. the 4 major comments →

arxiv 2510.20853 v2 pith:JSX7Z6DU submitted 2025-10-22 eess.AS cs.CLcs.SD

Beyond Hearing: Learning Task-Agnostic ExG Representations from Earphones via Physiology-Informed Tokenization

Hyungjun Yoon , Seungjoo Lee , Yu Yvonne Wu , Xiaomeng Chen , Taiting Lu , Freddy Yifei Liu , Taeckyung Lee , Hyeongheon Cha
show 6 more authors
Haochen Zhao Gaoteng Zhao Dongyao Chen Cecilia Mascolo Sung-Ju Lee Lili Qiu
This is my paper
classification eess.AS cs.CLcs.SD
keywords physiology-informed multi-band tokenizationtask-agnostic ExG representationearphone-based electrophysiologyself-supervised pre-trainingreconstruction objectiveDailySense datasetfive human sensesbiomedical signal processing
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.

The reading

This paper tries to establish that electrophysiological (ExG) signals recorded from earphones can be turned into task-agnostic representations by splitting them into a fixed set of 12 physiology-informed frequency sub-bands — EEG delta/theta/alpha/beta/gamma, EMG low/mid/high, EOG, ECG low/high, and QRS — and then pre-training a bidirectional state-space encoder with six reconstruction objectives on 50 hours of unlabeled free-living data. The central claim is that this decomposition, plus reconstruction pre-training, removes the need for task-specific filters or architectures: one model, fine-tuned per task, outperforms specialized baselines on gaze tracking, video/audio interest, taste, touch, smell, emotion recognition, sleep staging, and motor imagery. If correct, it matters because earphone-based sensing is cheap, unobtrusive, and scalable, making everyday physiological monitoring and ExG foundation models practical. The paper also contributes the hardware prototype and DailySense, the first ExG benchmark spanning the five human senses.

Core claim

At the core is Physiology-informed Multi-band Tokenization (PiMT): instead of feeding raw wide-band ExG or a task-selected narrow band, every channel is passed through a fixed bank of 12 canonical sub-band filters, each tied to a known physiological modality. Each sub-band is patched and embedded, with tokens ordered frequency-first, channel-second, time-last, and processed by a bidirectional state-space encoder. Pre-training on unlabeled free-living ear-ExG uses six complementary reconstruction tasks — autoencoding, masked reconstruction, and amplitude/phase reconstruction in both original and masked forms — which jointly force the encoder to retain temporal, spectral, and contextual inform

What carries the argument

The load-bearing object is the physiology-informed filter bank: 12 fixed sub-band filters spanning EEG delta/theta/alpha/beta/gamma, EMG low/mid/high frequency, EOG overall, ECG low/high frequency, and the QRS complex. Decomposing every channel into these bands and tokenizing patches in a frequency-first ordering gives the encoder fine-grained, task-independent access to spectral content; the reconstruction pre-training (six objectives with decoders) makes the representations usable without labels. The bidirectional state-space backbone keeps sequence modeling linear-time despite the added frequency dimension.

Load-bearing premise

The DailySense sensory labels (sweet vs. sour, floral vs. sour, rough vs. smooth) are assumed to reflect the intended perceptual state, but the protocols differ in mouth, face, and sniffing muscle activity that ear electrodes can record as EMG — the paper's own appendix acknowledges auricular electrodes capture facial-expression EMG — so without session- or subject-excluded validation, the strong within-session accuracy may partly reflect muscle-artifact classification.

What would settle it

Run the taste, smell, and touch protocols with sham conditions — sipping a neutral liquid without taste, sniffing an empty bag, rubbing the same texture while blindfolded — and check whether a model trained on the true labels can still separate the classes; alternatively, evaluate with a leave-one-session-out split where each recording session is held out, since the paper's own cross-subject F1 drops to 0.586.

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

If this is right

  • A single pre-trained ExG encoder can be fine-tuned to tasks as different as gaze estimation, taste/smell/touch classification, emotion recognition, sleep staging, and motor imagery without changing the tokenizer or choosing a frequency band.
  • The multi-band decomposition alone (without pre-training) already improves over wide-band or single-band baselines; adding free-living pre-training yields further gains, showing unlabeled in-the-wild data is useful.
  • Earphone electrodes capture sufficient EEG, EOG, and EMG information for these tasks, making a low-cost earhook form factor a viable alternative to lab headsets for everyday sensing.
  • Reconstruction-based pre-training scales with data: larger free-living corpora reduce reconstruction loss and continue improving downstream gaze regression even after classification performance saturates.
  • On-device inference is real-time on a smartphone (25 ms latency, 266 MB memory), so the approach is deployable in a wearable companion-device setting.

Where Pith is reading between the lines

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

  • If this holds, earphone-based ExG could become a general-purpose physiological sensing layer for context-aware wearables — detecting user state across senses with one model — but only if the sensory-classification results survive session- and subject-excluded validation, which the paper's own cross-subject drop (average F1 0.586) leaves open.
  • The tokenization is encoder-agnostic in spirit; the paper shows a state-space backbone beats a Transformer with the same multi-band tokens, but the band decomposition itself might transfer to other architectures or to larger foundation-model scaling regimes.
  • A direct test of the physiological claim would be to run the taste, smell, and touch protocols with sham stimuli (same mouth/nose movements but no actual taste/smell/texture difference); if a model still separates classes, the labels are carried by muscle activity rather than perception.
  • The 12-band boundaries are grounded in textbook physiology but are still arbitrary; a learned or data-driven band allocation would test whether the fixed physiological prior is the active ingredient or whether finer spectral resolution alone drives the gains.
Share X Bluesky LinkedIn Reddit HN

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 / 6 minor

Summary. The paper presents NeuroBuds, an earphone-based ExG sensing prototype, and DailySense, a dataset with 50 hours of free-living and 20 hours of labeled ExG recordings spanning sight, hearing, taste, touch, and smell. It introduces PiMT, which decomposes ExG into 12 physiology-informed sub-band tokens, encodes them with a Bidirectional-Mamba backbone, and pre-trains with six reconstruction objectives on unlabeled free-living data. The authors report an average F1 of 0.876 on DailySense and consistent improvements over PatchTST and Bidirectional-Mamba on DREAMER, SEED, Sleep-EDF, and BCI Competition IV 2b. The central claim is that PiMT learns task-agnostic ExG representations that generalize across diverse sensory and cognitive tasks.

Significance. The paper addresses a relevant problem and makes a concrete hardware/data contribution; conditional release of DailySense would be valuable. The public-benchmark results (Table 2) and the ablation tables provide useful evidence that multi-band tokenization plus reconstruction pre-training can transfer across datasets and tasks. The method is described in sufficient detail to reproduce, and the authors explicitly acknowledge limitations in Section 5.7 and Appendix L. However, the central 'task-agnostic five-sense' claim rests on a within-session evaluation with participant overlap (Section 5.1, Appendix F); the paper's own cross-subject numbers (Table 7) are much weaker. Because the conclusion is broader than the evidence currently supports, major revision is required.

major comments (4)
  1. [§5.1, Appendix F, Table 7] The headline DailySense result (avg F1 0.876) is obtained by randomly shuffling 4-second windows and applying an 80/20 split (Appendix F), so windows from the same participant/session appear in both training and test sets. Under a cross-subject split (Table 7), average F1 drops to 0.586, near chance for binary tasks, and PiMT no longer beats PatchTST on Video (0.612 vs 0.654) or Audio (0.578 vs 0.595). This indicates that the within-session numbers are substantially inflated by subject/session identity and/or task-correlated artifacts. The claim that PiMT learns task-agnostic representations across the five senses therefore needs a subject-disjoint, session-disjoint evaluation with artifact controls; Table 1 alone does not support it.
  2. [Appendix C, Appendix B, Table 3] The sensory protocols confound the target sensation with stereotyped motor behavior: taste involves sipping/holding chocolate milk vs vinegar; smell involves sniffing lavender vs vinegar; touch involves rubbing rough vs smooth materials. Appendix B states that the auricular electrodes capture facial-expression EMG, so mouth, face, and sniffing muscle activity is a natural confound. A model can classify these tasks from EMG/EOG artifacts without decoding the intended sensory state. Please provide artifact-baseline analyses (e.g., EMG-band-only classification, rest-period contrasts, or sham tasks) and/or an artifact-rejection pipeline; otherwise the 'five senses' interpretation of Table 1 is not established.
  3. [§5.6, Figure 7 vs Table 7] The text says that LOSO with the target subject excluded from pre-training achieves performance 'comparable to full pre-training,' but Table 7 reports a cross-subject average F1 of 0.586, far below 0.876. These statements are not reconciled. Please clarify the experimental setup for Figure 7 (is the target subject still present in fine-tuning? are the same task windows used?) and report the underlying numbers with error bars. As written, the two findings are difficult to reconcile and undermine the cross-subject generalization claim.
  4. [§3.2, §3.3, Appendix G, Table 5, Figure 4] The claim that PiMT is task-agnostic is weakened by the fact that several components (band ordering, patch size, λ weights, 12-band design) were selected based on performance on the same DailySense tasks. The saliency analysis then reads activation patterns back onto these tasks, so it does not provide independent confirmation of task generality. At minimum, please report which choices were made without access to DailySense labels and evaluate sensitivity to these choices on held-out tasks. The public benchmarks help, but Table 2 compares only two baselines, so the statement that PiMT 'consistently outperforms' is stronger than the evidence.
minor comments (6)
  1. [Figure 1 caption] The caption uses 'PiFT' in one place; this should be 'PiMT'.
  2. [§3.1] There is a duplicated word: 'tokens that that grant the model'.
  3. [§3.3] The notation for masked frequency reconstruction is inconsistent: the text writes 'producing ˆpMA and ˆpMP' but then uses 'ˆpA = D_MA' and 'ˆpP = D_MP'. Please define the masked targets consistently.
  4. [Appendix G vs §4] Appendix G says the representation model is trained on 'the entire 40 hours of free-living data,' while Section 4 and the abstract state 50 hours. Reconcile this discrepancy.
  5. [Appendix G] Typo: 'bi-directionanl mamba' should be 'bidirectional Mamba'.
  6. [Table 7] Cross-session standard deviations are very large (e.g., Video 0.697 ± 0.249, Audio 0.698 ± 0.188). This should be discussed, since it limits the strength of the cross-session robustness claim.

Circularity Check

0 steps flagged

No significant circularity: independent public benchmarks support the central claim; DailySense limitations are validity issues, not circular reasoning.

full rationale

PiMT's derivation chain is self-contained rather than circular. The input representation is a fixed 12-band filter bank defined from physiological references (Section 3.1); the encoder, reconstruction objectives, and fine-tuning are standard pipelines with no target label fed back into the definition of the method. The central generalization claim is anchored by four external benchmarks (DREAMER, SEED, Sleep-EDF, BCI Competition IV 2b, Table 2) with independently rerun baselines, so the conclusion does not rest on the authors' own prior result or on an identity between a fitted parameter and a reported prediction. The paper discloses the main threat to the DailySense headline: Appendix F's random 80/20 window shuffle can leak subject/session identity, and Table 7 reports cross-subject average F1 falling to 0.586, with PiMT losing to PatchTST on Video and Audio. That is a validity/leakage concern, not a circularity: no equation reduces a 'predicted' benchmark number to a fitted input by construction. Hyperparameter choices (λ values, patch size, token ordering) are described as empirically selected; if those choices were tuned on DailySense, the reported DailySense F1 is optimistically biased, but the same configuration is transferred to public datasets, which remain independent evidence. There is no load-bearing self-citation or imported uniqueness/ansatz chain. Therefore no circular step is identified.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 1 invented entities

The central claim rests on hand-selected physiology band definitions, unvalidated ear-electrode signal quality, a window-split independence assumption, and the assumption that sensory labels are not confounded by task-correlated muscle activity. These are domain assumptions rather than derived facts; the paper provides only partial support for each (visualizations, model performance, acknowledged cross-user drops). NeuroBuds is a new device with no independent validation; the 12-band tokenizer and hyperparameters are manually chosen.

free parameters (6)
  • 12-band filter bank edges = See §3.1: EEG δ0.5–4/θ4–8/α8–13/β13–30/γ30–100; EMG 15–45/45–95/95–100; EOG 0.1–20; ECG; QRS 8–50
    Freely chosen canonical bands, central to PiMT; no formal selection; several bands overlap (EEG γ 30–100 vs EMG-HF 95–100).
  • lambda pretraining loss weights = AE=2, MR=2, others=1
    Set heuristically in Appendix G after monitoring convergence; no sweep or validation.
  • patch size = 0.5 s
    Selected empirically in Table 5 for best average task performance.
  • token scanning order = frequency-first, channel-second, time-last
    Section 3.2 says chosen 'based on empirical validation'; ordering affects sequence structure.
  • masking ratio = 50%
    Fixed during pretraining (Appendix G); no reported sensitivity analysis.
  • Mamba depth = 8 (Appendix F) vs 16 (Appendix G) layers
    Manual architecture choice; inconsistent reporting hampers replication.
axioms (4)
  • domain assumption Canonical physiological bands from §3.1 are a sufficient decomposition for all relevant ExG task features.
    If the fixed 12-band set omits or misplaces a task-relevant band, PiMT cannot recover it; no search over band plans is reported.
  • domain assumption Earphone electrodes (T7–T10, FT7–FT10, TP7–TP10 plus auricular/lateral electrodes) record genuine EEG/EOG/EMG with adequate SNR.
    Appendix B validates signal quality only through the model's own performance and visual inspection, not against reference electrodes.
  • domain assumption Four-second, randomly split windows are independent and the 80/20 per-participant split measures generalization.
    Same-session windows are likely autocorrelated; Table 7's cross-session/cross-subject drops suggest non-independence inflates Table 1 scores.
  • domain assumption Task labels correspond to the intended sensory modality and not to correlated motor/expression artifacts.
    Tasting, smelling, and touching involve facial/jaw/sniff muscle activity; ear EMG can encode these confounds; no control task or artifact ablation separates them.
invented entities (1)
  • NeuroBuds earphone ExG sensing platform no independent evidence
    purpose: Unobtrusive collection of free-living and task-specific ExG data from around-the-ear electrodes
    Described in Appendix A with custom PCB, but no third-party or clinical-grade validation of signal quality; its claimed sensing capabilities are supported only by the paper's own recordings.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Hearing: Learning Task-Agnostic ExG Representations from Earphones via Physiology-Informed Tokenization." pith.science (2026). https://pith.science/paper/JSX7Z6DU

@misc{pith2026251020853,
  author       = {Pith},
  title        = {Pith review of: Beyond Hearing: Learning Task-Agnostic ExG Representations from Earphones via Physiology-Informed Tokenization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JSX7Z6DU}},
  note         = {Machine review of arXiv:2510.20853}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Electrophysiological (ExG) signals offer valuable insights into human physiology, yet building foundation models that generalize across everyday tasks remains challenging due to two key limitations: (i)~insufficient data diversity, as most ExG recordings are collected in controlled labs with bulky, expensive devices; and (ii)~task-specific model designs that require tailored processing (i.e., targeted frequency filters) and architectures, which limit generalization across tasks. To address these challenges, we introduce an approach for scalable, task-agnostic ExG monitoring in the wild. We collected 50 hours of unobtrusive free-living ExG data with an earphone-based hardware prototype to narrow the data diversity gap. At the core of our approach is Physiology-informed Multi-band Tokenization (PiMT), which decomposes ExG signals into 12 physiology-informed tokens, followed by a reconstruction task to learn robust representations. This enables adaptive feature recognition across the full frequency spectrum while capturing task-relevant information. Experiments on our new DailySense dataset, the first to enable ExG-based analysis across five human senses, together with four public ExG benchmarks, demonstrate that PiMT consistently outperforms state-of-the-art methods across diverse tasks.

Figures

Figures reproduced from arXiv: 2510.20853 by Cecilia Mascolo, Dongyao Chen, Freddy Yifei Liu, Gaoteng Zhao, Haochen Zhao, Hyeongheon Cha, Hyungjun Yoon, Lili Qiu, Seungjoo Lee, Sung-Ju Lee, Taeckyung Lee, Taiting Lu, Xiaomeng Chen, Yu Yvonne Wu.

Figure 1
Figure 1. Figure 1: Overview of PiMT. ExG signals are decomposed into 12 sub-bands via Physiology [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of DailySense dataset. Using our earphone-based ExG analysis device, Neu￾roBuds, we collect free-living unlabeled data for task-agnostic pre-training and labeled data spanning five human senses, serving as benchmarks for downstream tasks. the PCB as shown in [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of different ExG tokenization strategies: 1-band (0.1 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Saliency analysis demonstrating how the model dynamically captures task-relevant fre [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Test loss across dif￾ferent pre-training data scales. 0 10 30 50 100 Data Scale(%) 0.800 0.825 0.850 0.875 0.900 F1 Score Classification ( ) 0 10 30 50 100 Data Scale(%) 6.0 6.1 A n g ula r E r r o r ( ) Regression ( ) [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: LOSO Performance when the target sub￾ject’s data is either included or excluded. Leave-One-Subject-Out evaluation. To fur￾ther assess cross-participant generalization, we conducted leave-one-subject-out (LOSO) exper￾iments on DailySense. In DailySense, some participants contributed to both free-living pre￾training data and task-specific data. Therefore, in LOSO, for each target user, their data was exclude… view at source ↗
Figure 8
Figure 8. Figure 8: Raw ExG signals from DailySense dataset and their decomposition into twelve physiology [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of Mamba and Transformer backbones. [PITH_FULL_IMAGE:figures/full_fig_p017_9.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

19 extracted references · 3 linked inside Pith

  1. [5]

    However, it benefits gaze regression, as the participant’s gaze is more likely to remain fixed within a shorter temporal window

    A smaller patch size provides less contextual information for each classification window, which may limit performance. However, it benefits gaze regression, as the participant’s gaze is more likely to remain fixed within a shorter temporal window. In contrast, larger patch sizes offer more temporal context for classification tasks but increase the likelih...

    2024

  2. [7]

    URLhttps://dx.doi.org/10.1088/1741-2552/aace8c

    doi: 10.1088/1741-2552/aace8c. URLhttps://dx.doi.org/10.1088/1741-2552/aace8c. Robert Leeb, Clemens Brunner, G Müller-Putz, A Schlögl, and GJGUOT Pfurtscheller. Bci competi- tion 2008–graz data set b.Graz University of Technology, Austria, 16:1–6,

    doi:10.1088/1741-2552/aace8c 2008

  3. [11]

    10471803

    doi: 10.1109/RIVF60135.2023. 10471803. Christopher Wang, Vighnesh Subramaniam, Adam Uri Yaari, Gabriel Kreiman, Boris Katz, Ignacio Cases, and Andrei Barbu. Brainbert: Self-supervised representation learning for intracranial recordings. InThe Eleventh International Conference on Learning Representations,

    arXiv 2023

  4. [12]

    Xiuxin Xia, Yuchao Yang, Yan Shi, Wenbo Zheng, and Hong Men

    URL https://proceedings.neurips.cc/paper_files/paper/ 2024/file/3fe2a777282299ecb4f9e7ebb531f0ab-Paper-Conference.pdf. Xiuxin Xia, Yuchao Yang, Yan Shi, Wenbo Zheng, and Hong Men. Decoding taste information in human brain: A temporal and spatial reconstruction data augmentation method coupled with taste eeg,

    2024

  5. [13]

    Chaoqi Yang, M Westover, and Jimeng Sun

    URLhttps://arxiv.org/abs/2307.05365. Chaoqi Yang, M Westover, and Jimeng Sun. Biot: Biosignal transformer for cross-data learning in the wild.Advances in Neural Information Processing Systems, 36:78240–78260,

    Pith/arXiv arXiv

  6. [14]

    Brainwave: A brain signal foundation model for clinical applications.arXiv preprint arXiv:2402.10251,

    Zhizhang Yuan, Fanqi Shen, Meng Li, Yuguo Yu, Chenhao Tan, and Yang Yang. Brainwave: A brain signal foundation model for clinical applications.arXiv preprint arXiv:2402.10251,

    Pith/arXiv arXiv

  7. [16]

    Each side includes five electrodes: the top ones on the left and right act as bias and reference, while the remaining eight serve as signal channels

    13 A NEUROBUDSHARDWAREDESIGN To enable large-scale, in-the-wild ExG data collection, we built an earphone-based sensing platform consisting of two main components: Earphone-Shaped Sensing Array:To adopt a earhook-style form factor, We use a commercial earphone (Powerbeats PB123) as the backbone, and wrap conductive tape around the frame to form electrodes...

    2017

  8. [17]

    (model: Tobii 4C Eye Tracker). • Auditory and Video Interest Inference:Inspired by SEED and DREAMER datasets (Zheng & Lu, 2015; Katsigiannis & Ramzan, 2018), this experiment explored the correlation between ExG signals and engagement with visual/auditory stimuli. Partici- pants were asked to watch or listen to video clips. After each session, they rated t...

    2015

  9. [18]

    as our backbone architecture, which has demon- strated state-of-the-art performance across various time-series tasks (Zerveas et al., 2021; Song et al., 2022). To evaluate its effectiveness on ExG signals, we compare it against Transformer-based architecture, PatchTST (Nie et al., 2023), which showed strong performance in our main evaluation (Section 5.2)...

    2021

  10. [1996]

    12 Hoang-Thuy-Tien V o, Thi-Nhu-Quynh Nguyen, Do Duc Cuong, and Tuan Van Huynh

    doi: 10.1161/01.CIR.93.5.1043. 12 Hoang-Thuy-Tien V o, Thi-Nhu-Quynh Nguyen, Do Duc Cuong, and Tuan Van Huynh. Classification taste-eeg signals using base neural network. In2023 RIVF International Conference on Computing and Communication Technologies (RIVF), pp. 107–111,

    doi:10.1161/01.cir.93.5.1043

  11. [2000]

    Daniel R Kramer, Krista Lamorie-Foote, Michael Barbaro, Morgan B Lee, Terrance Peng, Angad Gogia, George Nune, Charles Y Liu, Spencer S Kellis, and Brian Lee

    doi: 10.1109/10.867928. Daniel R Kramer, Krista Lamorie-Foote, Michael Barbaro, Morgan B Lee, Terrance Peng, Angad Gogia, George Nune, Charles Y Liu, Spencer S Kellis, and Brian Lee. Utility and lower limits of frequency detection in surface electrode stimulation for somatosensory brain-computer interface in humans.Neurosurgical focus, 48(2):E2,

    doi:10.1109/10.867928

  12. [2006]

    Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805,

    Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805,

    Pith/arXiv arXiv

  13. [2015]

    Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang

    doi: 10.1109/TAMD.2015.2431497. Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: efficient visual representation learning with bidirectional state space model. In Proceedings of the 41st International Conference on Machine Learning, ICML’24. JMLR.org,

    arXiv 2015

  14. [2016]

    ISBN 9781450342636

    Association for Computing Machinery. ISBN 9781450342636. doi: 10.1145/2994551.2994562. URL https://doi.org/10.1145/2994551.2994562. Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers. InThe Eleventh International Confer- ence on Learning Representations,

    arXiv

  15. [2017]

    The standardized eeg electrode array of the ifcn.Clinical neurophysiology, 128(10):2070–2077,

    Margitta Seeck, Laurent Koessler, Thomas Bast, Frans Leijten, Christoph Michel, Christoph Baum- gartner, Bin He, and Sándor Beniczky. The standardized eeg electrode array of the ifcn.Clinical neurophysiology, 128(10):2070–2077,

    2070

  16. [2018]

    doi: 10.1109/JBHI.2017.2688239. B. Kemp, A.H. Zwinderman, B. Tuk, H.A.C. Kamphuisen, and J.J.L. Oberye. Analysis of a sleep- dependent neuronal feedback loop: the slow-wave microcontinuity of the eeg.IEEE Transactions on Biomedical Engineering, 47(9):1185–1194,

    arXiv 2017

  17. [2022]

    Andrea Apicella, Pasquale Arpaia, Giovanna Mastrati, and Nicola Moccaldi

    doi: 10.1109/ICEE55646.2022.9827406. Andrea Apicella, Pasquale Arpaia, Giovanna Mastrati, and Nicola Moccaldi. Eeg-based detection of emotional valence towards a reproducible measurement of emotions.Scientific Reports, 11(1): 21615,

    arXiv 2022

  18. [2023]

    Multimodal multi- view spectral-spatial-temporal masked autoencoder for self-supervised emotion recognition

    Pengxuan Gao, Tianyu Liu, Jia-Wen Liu, Bao-Liang Lu, and Wei-Long Zheng. Multimodal multi- view spectral-spatial-temporal masked autoencoder for self-supervised emotion recognition. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1926–1930. IEEE,

    2024

  19. [2024]

    doi: 10.1016/j.aei.2024.102522

    ISSN 1474-0346. doi: 10.1016/j.aei.2024.102522. URL https://doi.org/10.1016/j.aei.2024.102522. Iffat Farhana, Jungpil Shin, Shabbir Mahmood, Md Rabiul Islam, and Md Khademul Islam Molla. Emotion recognition using narrowband spatial features of electroencephalography.IEEE Access, 11:44019–44033,

    arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.