Pith. sign in

REVIEW 3 major objections 5 minor 42 references

One EEG model decodes three driver-brain states at once

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 ↗

DS-MTNet uses frequency-constrained source decomposition and reusable information slots to jointly decode three EEG-based cognitive readouts in a driving task, outperforming single-task and multi-task baselines.

T0 review reviewed 2026-07-08 challenge →

load-bearing objection Structured multi-task EEG decoder with solid experimental protocol but ablations that confound structure with capacity the 3 major comments →

arxiv 2607.06297 v1 pith:DPSOSHPQ submitted 2026-07-07 cs.HC

DS-MTNet:Structured Multi-Task EEG Decoding for Human-Machine Collaboration

Xinjia Yu , Yang Zhou , Jing Yang , Tielin Shi , Tao Cheng This is my paper
classification cs.HC
keywords ds-mtnetdecodingmulti-taskevidencereadoutscollaborationhmc-relevanthuman-machine
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 proposes DS-MTNet, a structured multi-task EEG decoding framework that simultaneously extracts three distinct operator-relevant readouts from a single EEG window: environmental-event processing (lane-departure detection), response preparation (steering-response stage), and internal alertness state (reaction-time-defined). The central design claim is that decomposing mixed EEG signals into frequency-constrained candidate sources, organizing multi-view evidence into reusable information slots, and using task-specific gated readouts yields better joint decoding than training separate models or using standard multi-task sharing. The architecture has two parallel branches: a source branch that performs ICA-style spatial whitening, SincNet band-pass filtering, and band-wise unmixing to create a shared source dictionary, then routes task-specific source subsets via a softmax router; and an information branch that fuses waveform, source-embedding, and power views into reusable slots read by dual gating mechanisms (task-slot and task-slot-band gates). Evaluated on the SADT driving EEG dataset under 20-fold leave-one-subject-out cross-validation, DS-MTNet achieved the best mean accuracy and F1 across all three tasks compared to traditional, single-task deep, and multi-task baselines, with the most statistically robust gains for steering-response stage decoding (significant against all five representative baselines, q < 0.05). Ablation studies showed that removing reusable information slots caused the largest performance drop (4.17 ACC points), followed by removing the source view (2.33 points) and source decomposition (1.88 points), supporting the claim that the structured decomposition, not generic multi-task capacity, drives the improvement. Interpretability analyses indicated that each task predominantly activated a different information slot and that task-routed pathways exhibited distinct spectral profiles (significant Task x Band interaction, permutation p = 0.005).

Core claim

The paper's central discovery is that a structured decomposition of EEG evidence into frequency-constrained source components, reusable multi-view information slots, and task-specific gated readouts enables a single model to jointly decode three heterogeneous operator-relevant brain readouts with cross-subject generalization, and that this decomposition is functionally meaningful rather than incidental: each task preferentially activates a different information slot (70-73% diagonal dominance) and produces a distinct spectral profile, providing evidence that the model separates mixed neural evidence into task-appropriate pathways rather than learning a single shared representation.

What carries the argument

DS-MTNet's architecture centers on two mechanisms: (1) a source branch that decomposes EEG into a shared dictionary of frequency-band-constrained candidate sources (via SincNet filtering and band-wise unmixing) and routes task-specific subsets through a softmax task router, and (2) an information branch that fuses three views (waveform, source embeddings, power) into reusable information slots accessed by dual gates (task-slot gate selecting among slots, task-slot-band gate selecting frequency-band-specific power evidence). A non-negative logit fusion head combines source-branch and information-branch outputs per task. Three regularizers (reconstruction, correlation, sparsity) constrain the源

Load-bearing premise

The claim that performance gains stem from the structured decomposition (source routing plus reusable slots) rather than from generic multi-task capacity or dataset-specific overfitting rests on an ablation study that removes components one at a time without controlling for parameter count or training budget across variants, so a model with equivalent parameters but unstructured sharing could potentially match performance without the decomposition.

What would settle it

A parameter-matched, training-budget-matched multi-task baseline with unstructured sharing that achieves equivalent or better performance on all three tasks would undermine the central design claim that the structured decomposition (source routing, reusable slots, gated readouts) is the causal mechanism behind the performance gains.

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

If this is right

  • If the structured decomposition generalizes beyond driving, HMC systems in surgical assistance, industrial robotics, or autonomous driving could use a single EEG model to simultaneously monitor operator perception, response preparation, and alertness, replacing multiple separately trained decoders.
  • The task-slot allocation pattern (each task dominating a different slot) suggests that learned information slots may correspond to partially separable neural processes, which could be validated against independent neurophysiological evidence.
  • The frequency-gate's task-specific selectivity (almost entirely benefiting alertness-state decoding) implies that different HMC-relevant readouts may require different spectral resolution, informing how future EEG sensing systems allocate frequency-band processing.
  • The offline-only evaluation leaves open whether the decoded readouts can be integrated into real-time closed-loop HMC systems for adaptive warning or personalized assistance within latency constraints.
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

3 major / 5 minor

Summary. The paper proposes DS-MTNet, a structured multi-task EEG decoding framework for human-machine collaboration (HMC). The model decomposes EEG windows through a source branch (frequency-constrained candidate sources, task-specific routing) and an information branch (three views—waveform, source, power—mapped into reusable slots with dual gating). Three binary decoding tasks are defined on the SADT dataset: lane-departure detection, steering-response stage, and reaction-time-defined alertness state. The model is evaluated under 20-fold leave-one-subject-out (LOSO) cross-validation against traditional, single-task deep, and multi-task baselines, with FDR-corrected paired statistical tests, ablation studies, and interpretability analyses (permutation tests, task-slot gate inspection, model-guided spectral profiles). DS-MTNet achieves the best mean ACC/F1 on all three tasks, with the most robust gains on Task 2 (significant against all five representative baselines, q < 0.05).

Significance. The paper addresses a genuine gap: organizing multiple HMC-relevant EEG readouts within a unified representation framework, rather than treating each as an independent classification problem. The experimental protocol is rigorous—20-fold LOSO, FDR-corrected paired comparisons, and ablations across eight variants. The interpretability analyses include permutation-based Task × Band interaction tests and subject-level statistics, which is commendable. The model-guided spectral analysis (Fig. 4) is a nice touch, providing falsifiable, task-dependent spectral profiles. The work is primarily empirical and does not make unfalsifiable claims; the authors are appropriately cautious about Task 3's weaker statistical support and the descriptive nature of the source-router maps.

major comments (3)
  1. §V.C, Table III (InfoSlot row): The 'No information slots' ablation—the largest single drop (4.17 ACC points)—replaces the gated slot mechanism (Q=4 slots, task-specific gates, band-specific adaptation per Eq. 5) with direct concatenation of the three views followed by classification. This simultaneously changes (a) parameter count, (b) computational depth (slot transforms φ_q add a layer), (c) intermediate representation dimensionality, and (d) bottleneck properties. The ablation therefore cannot distinguish whether the slot structure per se is responsible, or whether any intermediate processing of comparable capacity would achieve similar gains. This is load-bearing because the central design claim is that 'performance was related to the structured shared-plus-specific decomposition, not only to the presence of multiple task heads' (§V.C). A parameter- and depth-matched unstructured ab
  2. §V.D, Fig. 3b: The task-slot gate is initialized so that each task starts from one slot (acknowledged in the text). The reported diagonal dominance (70.6%, 72.6%, 70.6%) could partly reflect this initialization prior rather than purely learned task separation. The paper's defense—that the equal-slot-weight ablation (Table III) reduces ACC by 1.62 points—shows the gate is functionally useful but does not rule out that initialization biases the interpretability pattern. A random initialization or a permutation test shuffling task-slot assignments would more directly test whether the diagonal structure is learned rather than inherited. This matters because Fig. 3b is presented as the 'main and statistically supported evidence for task-specific slot allocation.'
  3. §IV.A, Task definitions: Task 2 (steering-response stage) uses preparation epochs [-3, 0] s and action-onset epochs [-1, +2] s around response onset. These windows overlap in the [-1, 0] s interval. Could the same EEG segment appear in both classes, or are the epoch sets disjoint by construction? If there is temporal overlap, the classification boundary may partly reflect time-since-response-onset rather than response-stage-related neural activity. This should be clarified, as Task 2 is the task where DS-MTNet shows its most robust gains.
minor comments (5)
  1. §III.B, Eq. (3): The notation α_{m,r} is defined as a softmax over a_{m,r}, but the indices b,k appear in the sum without being explicitly tied to the softmax dimension. A brief clarification that softmax is over all F×K candidates would help.
  2. Table I: EEG Conformer performs notably worse than other deep baselines on Tasks 1 and 2 (69.63% and 67.62% ACC). A brief comment on why would help readers assess whether this reflects a known limitation or a setup issue.
  3. §V.E, Fig. 4: Only Task 1 reaches significance against zero in the band-wise contrasts. The text is appropriately cautious, but the phrase 'task-appropriate way' for Tasks 2 and 3 may overstate the directional evidence given non-significance.
  4. References [28] and [30] cite 2026 dates; please verify these are not preprints with incorrect metadata.
  5. §III.D, Eq. (7): The reconstruction loss L_rec reconstructs X_w from S, but the relationship between the reconstruction target and the supervised objective is not discussed. A sentence on why reconstruction helps (beyond 'retaining signal information') would strengthen the motivation.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive review. The three major comments are well-taken and address genuine methodological concerns. We respond to each below and describe the revisions we will make.

read point-by-point responses

  1. Referee: §V.C, Table III (InfoSlot row): The 'No information slots' ablation confounds slot structure with parameter count, depth, dimensionality, and bottleneck properties. A parameter- and depth-matched unstructured ablation is needed to isolate the effect of slot structure per se.

    Authors: The referee is correct. The current InfoSlot ablation replaces the gated slot mechanism with direct concatenation of the three views, which simultaneously changes capacity, depth, dimensionality, and bottleneck structure. This confound means the ablation cannot distinguish whether the slot structure itself is responsible for the 4.17-point ACC drop, or whether any intermediate processing of comparable capacity would achieve similar gains. We agree this is a load-bearing concern because the central design claim is that structured shared-plus-specific decomposition—not generic capacity—drives performance. In the revision we will add a parameter- and depth-matched unstructured ablation: a variant in which the three views are processed by an MLP block with matched parameter count and depth to the slot transforms (φ_q), producing an intermediate representation of the same dimensionality, followed by the same task-specific classification head. This will isolate the contribution of the slot structure from generic capacity. If the matched unstructured variant performs comparably to the full model, we will revise our claim accordingly; if it performs worse, the result will strengthen the structured-decomposition argument. We will also soften the current wording from 'the drop mainly reflected the value of organizing multi-view EEG into reusable slots' to acknowledge that the original ablation could not fully separate structure from capacity. revision: yes

  2. Referee: §V.D, Fig. 3b: The diagonal dominance in task-slot gates could reflect initialization prior rather than learned task separation. A random initialization or permutation test would more directly test whether the diagonal structure is learned.

    Authors: This is a fair and important point. The task-slot gate is initialized so that each task starts from one slot, and the current defense (the equal-slot-weight ablation showing a 1.62-point ACC drop) demonstrates that the gate is functionally useful but does not rule out that the diagonal pattern is inherited from initialization rather than learned. We will address this in two ways. First, we will add a random-initialization variant (task-slot gates initialized uniformly) and report whether the diagonal structure still emerges after training. Second, we will add a permutation test: for each fold, we will shuffle task-slot assignments and compare the resulting gate patterns to the observed ones, providing a null distribution against which the diagonal dominance can be assessed. We will revise the text to explicitly acknowledge the initialization concern and to present the new results as the primary evidence for learned task-specific allocation. If the diagonal structure does not survive random initialization, we will revise the interpretability claim to state that the model exploits but does not necessarily discover task-specific slot allocation. revision: yes

  3. Referee: §IV.A, Task definitions: Task 2 uses preparation epochs [-3, 0] s and action-onset epochs [-1, +2] s, which overlap in [-1, 0] s. Could the same EEG segment appear in both classes?

    Authors: We thank the referee for catching this. The two epoch windows do overlap in the [-1, 0] s interval. In the current implementation, epochs are extracted independently for each class, so the same EEG samples in the [-1, 0] s interval can appear in both the preparation (negative) and action-onset (positive) classes. This means the classification boundary for Task 2 may partly reflect time-since-response-onset rather than purely response-stage-related neural activity. This is a genuine limitation. In the revision we will: (1) clarify in §IV.A that the epoch sets are not disjoint by construction and that temporal overlap exists; (2) add a non-overlapping variant where preparation epochs are restricted to [-3, -1] s (removing the overlap) and re-evaluate Task 2 performance under this stricter definition; and (3) discuss the implications for interpreting Task 2 results, noting that the robust gains on Task 2 should be understood in light of this epoch-definition issue. We will report both the original and non-overlapping results so that readers can assess the sensitivity of the findings to epoch construction. revision: yes

Circularity Check

0 steps flagged

No circularity found: the paper is an empirical ML study with held-out evaluation and no self-citation chain or definitional reduction.

full rationale

DS-MTNet is a standard empirical deep-learning paper. It proposes an architecture, trains it on labeled EEG data, and evaluates on held-out subjects via 20-fold leave-one-subject-out cross-validation. The 'predictions' are classification outputs on test data, not fitted constants re-emerging from definitions. No parameter is fit to a subset of data and then 'predicted' on closely related data. The ablation study (Table III) removes components and re-evaluates under the same protocol—this is a standard sensitivity analysis, not a circular derivation. The interpretability analyses (Fig. 3, Fig. 4) are descriptive and the paper explicitly acknowledges this: 'these profiles should be interpreted as model-guided descriptive evidence instead of direct physiological markers.' The task-slot gate initialization (each task starts from one slot) is acknowledged as a potential confound, and the paper argues against it being the sole explanation by citing the equal-slot-weight ablation. No self-citation chain is load-bearing: the architecture components (SincNet, EEGNet, ICA-style preprocessing, CGC, MMoE) are all attributed to external works. The design claim—that performance stems from structured decomposition rather than generic capacity—is supported by ablation evidence, though the skeptic's concern about uncontrolled parameter count is a validity threat, not a circularity issue. The derivation chain from architecture definition to trained model to held-out evaluation is self-contained and does not reduce to its inputs by construction.

Axiom & Free-Parameter Ledger

11 free parameters · 5 axioms · 3 invented entities

The model introduces 11+ categories of learnable parameters, most of which are standard neural network weights. The three invented entities (source dictionary, information slots, task-routed sources) are architectural constructs validated only internally via ablation and interpretability analysis. No external neurophysiological validation (e.g., comparison with ICA components, source localization, or independent EEG datasets) is provided. The domain assumptions are reasonable but unverified beyond this single dataset.

free parameters (11)
  • W_w (spatial whitening matrix) = learned, initialized from training-fold covariance
    C×C learnable matrix for channel conditioning before source separation
  • SincNet filter cutoffs = initialized to 1-4, 4-8, 8-13, 13-30, 30-100 Hz; learnable
    F=5 band-pass filters with learnable cutoffs under lower-bound constraints
  • W_b (band-wise unmixing matrices) = learned
    K×C matrices for each of 5 bands, producing K=8 candidate sources per band
  • a_{m,r} (task router logits) = learned
    Softmax routing weights over F×K=40 candidate sources for each task-slot pair
  • u_q (slot-view gate logits) = learned
    Q=4 slots, each with softmax over 3 views
  • η_{m,q} (task-slot-band gate) = learned
    F-dimensional band selection per task-slot pair
  • ρ_m (task-slot gate) = learned
    Q-dimensional slot selection per task
  • r_m (fusion gate scalar) = learned, softplus-constrained
    Per-task scalar weighting information branch contribution
  • λ_rec, λ_corr, λ_sparse = 0.005, 0.005, 0.001
    Regularization weights for reconstruction, correlation, and sparsity losses
  • Learning rate, weight decay, batch size = 0.001, 5e-5, 64
    Standard training hyperparameters
  • Task 3 RT threshold = top/bottom 25% per session
    Alertness label definition based on reaction time quartiles
axioms (5)
  • domain assumption EEG windows contain mixed but separable evidence for event-related, response-related, and state-related processes
    Invoked in §I and §II.B as motivation for the multi-task decomposition; not independently verified but consistent with EEG literature
  • domain assumption Frequency-band-limited source decomposition yields more interpretable and task-relevant components than raw channel-space features
    Underlies the SincNet + band-wise unmixing design in §III.B; supported by ablation but assumed a priori
  • ad hoc to paper Reusable information slots learned from data will align with task-relevant decompositions
    The slot-view gate and task-slot gate design (§III.C, Eq. 4-5) assumes slots will specialize; the initialization biases this toward diagonal structure
  • domain assumption The three defined tasks (lane departure, steering stage, RT alertness) are representative HMC-relevant readouts
    Stated in §I and §IV.A; reasonable but limited to one driving dataset
  • domain assumption Cross-subject generalization within one dataset generalizes to broader HMC scenarios
    Implicit in the conclusion (§VI); the paper acknowledges this as future work
invented entities (3)
  • Candidate source dictionary S (frequency-constrained) independent evidence
    purpose: Shared intermediate representation decomposing EEG into band-limited temporal sources
    The ablation (Table III, SrcDec variant) shows removing learned decomposition reduces performance by 1.88 ACC points, and Fig. 4 shows task-dependent spectral profiles. However, the sources are not validated against independent ICA or neurophysiological source localization.
  • Reusable information slots Z_q independent evidence
    purpose: Shared multi-view information components read by task-specific gates
    The ablation (Table III, InfoSlot variant) shows removing slots reduces performance by 4.17 ACC points, and Fig. 3b shows task-specific slot allocation. But slot contents are not independently validated against known neurophysiological constructs.
  • Task-routed source slots S̃_m independent evidence
    purpose: Task-specific soft mixtures over the candidate source dictionary
    The ablation (Table III, SrcRoute variant) shows shared routing reduces performance by 1.35 ACC points, and Fig. 3a shows task-dependent band preferences with FDR-corrected significance.

reviewed 2026-07-08 · how reviews work

0 comments
Cite this review

Pith. "Pith review of DS-MTNet:Structured Multi-Task EEG Decoding for Human-Machine Collaboration." pith.science (2026). https://pith.science/paper/DPSOSHPQ

@misc{pith2026260706297,
  author       = {Pith},
  title        = {Pith review of: DS-MTNet:Structured Multi-Task EEG Decoding for Human-Machine Collaboration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DPSOSHPQ}},
  note         = {Machine review of arXiv:2607.06297}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Current human-machine collaboration (HMC) systems rely on environment-facing sensors to observe visible actions and scene states, but the internal perceptual, intention-related, and state-related processes of operators remain insufficiently integrated into machine perception. Electroencephalography (EEG) provides a non-invasive, time-resolved modality to capture neural activity associated with these processes and can serve as an additional sensing channel in HMC. However, HMC-relevant EEG evidence is often mixed in continuous recordings. Existing EEG decoding methods usually target task-specific classification or aggregate prediction, so multiple HMC-relevant readouts are rarely organized in a unified EEG representation. To address this gap, this paper proposed the Decomposed-Source Multi-Task Network (DS-MTNet), a structured multi-task EEG decoding framework. DS-MTNet integrated three streams, namely EEG waveforms, task-routed source embeddings, and temporal-spectral power features, into reusable slots and used dual gating mechanisms to route task-specific components. The model was tested on a sustained-attention driving EEG dataset with three representative readouts: lane-departure-related epochs for environmental-event processing, steering-response stage for response preparation, and reaction-time-defined alertness state for internal state. DS-MTNet achieved the best mean performance among traditional, single-task deep, and multi-task EEG baselines, with the most robust gains observed for steering-response stage decoding. Ablation and interpretability analyses suggested that DS-MTNet jointly decoded multiple readouts and organized event-related, response-related, and state-related EEG evidence in a unified source-slot representation. These findings provide a computational step toward incorporating operator-related neural evidence into machine perception in HMC.

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

42 extracted references · 42 canonical work pages · 3 internal anchors

  1. [1]

    A reinforcement learning method for human -robot collaboration in assembly tasks,

    R. Zhang, Q. Lv, J. Li, J. Bao, T. Liu, and S. Liu, “A reinforcement learning method for human -robot collaboration in assembly tasks,” Rob. Comput. Integr. Manuf. , vol. 73, p. 102227, Feb. 2022, doi: 10.1016/j.rcim.2021.102227

    work page doi:10.1016/j.rcim.2021.102227 2022

  2. [2]

    Human –robot collaboration in industrial environments: A literature review on non - destructive disassembly,

    S. Hjorth and D. Chrysostomou, “Human –robot collaboration in industrial environments: A literature review on non - destructive disassembly,” Robotics and Computer -Integrated Manufacturing, vol. 73, p. 102208, Feb. 2022, doi: 10.1016/j.rcim.2021.102208

    work page doi:10.1016/j.rcim.2021.102208 2022

  3. [3]

    Toward human-vehicle collaboration for automated vehicles: A review and perspective,

    T. Huang, R. Fu, Q. Sun, Z. Deng, S. Huang, and L. Jin, “Toward human-vehicle collaboration for automated vehicles: A review and perspective,” IEEE Transactions on Intelligent Transportation Systems , vol. 26, no. 5, pp. 5649 –5673, May 2025, doi: 10.1109/TITS.2025.3530143

    work page doi:10.1109/tits.2025.3530143 2025

  4. [4]

    Review of human– robot collaboration in robotic surgery,

    Q. Wan, Y. Shi, X. Xiao, X. Li, and H. Mo, “Review of human– robot collaboration in robotic surgery,” Advanced Intelligent Systems, vol. 7, no. 2, p. 2400319, 2025, doi: 10.1002/aisy.202400319

    work page doi:10.1002/aisy.202400319 2025

  5. [5]

    A review of human –machine cooperation in the robotics domain,

    C. Yang, Y. Zhu, and Y. Chen, “A review of human –machine cooperation in the robotics domain,” IEEE Transactions on Human-Machine Systems, vol. 52, no. 1, pp. 12–25, Feb. 2022, doi: 10.1109/THMS.2021.3131684

    work page doi:10.1109/thms.2021.3131684 2022

  6. [6]

    Human –robot collaboration and machine learning: A systematic review of recent research,

    F. Semeraro, A. Griffiths, and A. Cangelosi, “Human –robot collaboration and machine learning: A systematic review of recent research,” Robotics and Computer -Integrated Manufacturing, vol. 79, p. 102432, Feb. 2023, doi: 10.1016/j.rcim.2022.102432

    work page doi:10.1016/j.rcim.2022.102432 2023

  7. [7]

    Feasibility of Detecting Potential Emergencies in Symbiotic Human-Robot Collaboration with a mobile EEG,

    A. Buerkle, T. Bamber, N. Lohse, and P. Ferreira, “Feasibility of Detecting Potential Emergencies in Symbiotic Human-Robot Collaboration with a mobile EEG,” Rob. Comput. Integr. Manuf., vol. 72, p. 102179, Dec. 2021, doi: 10.1016/j.rcim.2021.102179

    work page doi:10.1016/j.rcim.2021.102179 2021

  8. [8]

    Robotic vision for human-robot interaction and collaboration: A survey and systematic review,

    N. Robinson, B. Tidd, D. Campbell, D. Kulić, and P. Corke, “Robotic vision for human-robot interaction and collaboration: A survey and systematic review,” J. Hum.-Robot Interact., vol. 12, no. 1, p. 12:1-12:66, Feb. 2023, doi: 10.1145/3570731

    work page doi:10.1145/3570731 2023

  9. [9]

    Multimodal human – robot interaction for human -centric smart manufacturing: A survey,

    T. Wang, P. Zheng, S. Li, and L. Wang, “Multimodal human – robot interaction for human -centric smart manufacturing: A survey,” Advanced Intelligent Systems, vol. 6, no. 3, p. 2300359, 2024, doi: 10.1002/aisy.202300359

    work page doi:10.1002/aisy.202300359 2024

  10. [10]

    Safe human –robot collaboration for industrial settings: A survey,

    W. Li, Y. Hu, Y. Zhou, and D. T. Pham, “Safe human –robot collaboration for industrial settings: A survey,” J Intell Manuf, vol. 35, no. 5, pp. 2235–2261, Jun. 2024, doi: 10.1007/s10845- 023-02159-4

    work page doi:10.1007/s10845- 2024

  11. [11]

    Neural decoding of EEG signals with machine learning: A systematic review,

    M. Saeidi et al., “Neural decoding of EEG signals with machine learning: A systematic review,” Brain Sci, vol. 11, no. 11, p. 1525, Nov. 2021, doi: 10.3390/brainsci11111525

    work page doi:10.3390/brainsci11111525 2021

  12. [12]

    Predicting collision cases at unsignalized intersections using EEG metrics and driving simulator platform,

    X. Zhang and X. Yan, “Predicting collision cases at unsignalized intersections using EEG metrics and driving simulator platform,” Accident Analysis & Prevention, vol. 180, p. 106910, Feb. 2023, doi: 10.1016/j.aap.2022.106910

    work page doi:10.1016/j.aap.2022.106910 2023

  13. [13]

    Hybrid recurrent neural network architecture -based intention recognition for human–robot collaboration,

    X. Gao, L. Yan, G. Wang, and C. Gerada, “Hybrid recurrent neural network architecture -based intention recognition for human–robot collaboration,” IEEE Transactions on Cybernetics, vol. 53, no. 3, pp. 1578 –1586, Mar. 2023, doi: 10.1109/TCYB.2021.3106543

    work page doi:10.1109/tcyb.2021.3106543 2023

  14. [14]

    Augmented Recognition of Distracted Driving State Based on Electrophysiological Analysis of Brain Network,

    G. Qi, R. Liu, G. Wei, and A. Huang, “Augmented Recognition of Distracted Driving State Based on Electrophysiological Analysis of Brain Network,” Cyborg and Bionic Systems , vol. 5, Jul. 2024, doi: 10.34133/cbsystems.0130

    work page doi:10.34133/cbsystems.0130 2024

  15. [15]

    Sensory target detection at local and global timescales reveals a hierarchy of supramodal dynamics in the human cortex,

    M. Niedernhuber, F. Raimondo, J. D. Sitt, and T. A. Bekinschtein, “Sensory target detection at local and global timescales reveals a hierarchy of supramodal dynamics in the human cortex,” J. Neurosci., vol. 42, no. 46, pp. 8729 –8741, Nov. 2022, doi: 10.1523/JNEUROSCI.0658-22.2022

    work page doi:10.1523/jneurosci.0658-22.2022 2022

  16. [16]

    Electrophysiological underpinnings of reward processing: Are we exploiting the full potential of EEG?,

    G. M. Meyer, J. Marco -Pallaré s, P. Boulinguez, and G. Sescousse, “Electrophysiological underpinnings of reward processing: Are we exploiting the full potential of EEG?,” NeuroImage, vol. 242, p. 118478, Nov. 2021, doi: 10.1016/j.neuroimage.2021.118478

    work page doi:10.1016/j.neuroimage.2021.118478 2021

  17. [17]

    Toward reliable signals decoding for electroencephalogram: A benchmark study to EEGNeX,

    X. Chen, X. Teng, H. Chen, Y. Pan, and P. Geyer, “Toward reliable signals decoding for electroencephalogram: A benchmark study to EEGNeX,” Biomedical Signal Processing and Control , vol. 87, p. 105475, Jan. 2024, doi: 10.1016/j.bspc.2023.105475

    work page doi:10.1016/j.bspc.2023.105475 2024

  18. [18]

    Physics - informed attention temporal convolutional network for EEG - based motor imagery classification,

    H. Altaheri, G. Muhammad, and M. Alsulaiman, “Physics - informed attention temporal convolutional network for EEG - based motor imagery classification,” IEEE Transactions on Industrial Informatics, vol. 19, no. 2, pp. 2249–2258, Feb. 2023, doi: 10.1109/TII.2022.3197419

    work page doi:10.1109/tii.2022.3197419 2023

  19. [19]

    SpeechBrain-MOABB: An open-source python library for benchmarking deep neural networks applied to EEG signals,

    D. Borra, F. Paissan, and M. Ravanelli, “SpeechBrain-MOABB: An open-source python library for benchmarking deep neural networks applied to EEG signals,” Computers in Biology and Medicine, vol. 182, p. 109097, Nov. 2024, doi: 10.1016/j.compbiomed.2024.109097

    work page doi:10.1016/j.compbiomed.2024.109097 2024

  20. [20]

    Multi - channel EEG recordings during a sustained -attention driving task,

    Z. Cao, C. -H. Chuang, J. -K. King, and C. -T. Lin, “Multi - channel EEG recordings during a sustained -attention driving task,” Sci. Data , vol. 6, no. 1, p. 19, Apr. 2019, doi: 10.1038/s41597-019-0027-4

    work page doi:10.1038/s41597-019-0027-4 2019

  21. [21]

    Perception and sensing for autonomous vehicles under adverse weather conditions: A survey,

    Y. Zhang, A. Carballo, H. Yang, and K. Takeda, “Perception and sensing for autonomous vehicles under adverse weather conditions: A survey,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 196, pp. 146 –177, Feb. 2023, doi: 10.1016/j.isprsjprs.2022.12.021

    work page doi:10.1016/j.isprsjprs.2022.12.021 2023

  22. [22]

    Sensor -enabled safety systems for human – robot collaboration: A review,

    C. Scholz et al. , “Sensor -enabled safety systems for human – robot collaboration: A review,” IEEE Sensors Journal, vol. 25, no. 1, pp. 65–88, Jan. 2025, doi: 10.1109/JSEN.2024.3496905

    work page doi:10.1109/jsen.2024.3496905 2025

  23. [23]

    Sensor -based human –robot collaboration for industrial tasks,

    A. Angleraud, A. Ekrekli, K. Samarawickrama, G. Sharma, and R. Pieters, “Sensor -based human –robot collaboration for industrial tasks,” Robotics and Computer -Integrated Manufacturing, vol. 86, p. 102663, Apr. 2024, doi: 10.1016/j.rcim.2023.102663

    work page doi:10.1016/j.rcim.2023.102663 2024

  24. [24]

    Gaze-aware hand gesture recognition for intelligent construction,

    X. Wang, D. Veeramani, and Z. Zhu, “Gaze-aware hand gesture recognition for intelligent construction,” Engineering Applications of Artificial Intelligence, vol. 123, p. 106179, Aug. 2023, doi: 10.1016/j.engappai.2023.106179

    work page doi:10.1016/j.engappai.2023.106179 2023

  25. [25]

    Proactive human –robot collaboration: Mutual - cognitive, predictable, and self -organising perspectives,

    S. Li et al. , “Proactive human –robot collaboration: Mutual - cognitive, predictable, and self -organising perspectives,” Robotics and Computer -Integrated Manufacturing, vol. 81, p. 102510, Jun. 2023, doi: 10.1016/j.rcim.2022.102510

    work page doi:10.1016/j.rcim.2022.102510 2023

  26. [26]

    An adaptive human sensor framework for human–robot collaboration,

    A. Buerkle, H. Matharu, A. Al -Yacoub, N. Lohse, T. Bamber, and P. Ferreira, “An adaptive human sensor framework for human–robot collaboration,” Int. J. Adv. Manuf. Technol., vol. 119, no. 1, pp. 1233 –1248, Mar. 2022, doi: 10.1007/s00170 - 021-08299-2

    work page doi:10.1007/s00170 2022

  27. [27]

    Anomaly Detection Methods in Autonomous Robotic Missions,

    S. Chirayil Nandakumar, D. Mitchell, M. S. Erden, D. Flynn, and T. Lim, “Anomaly Detection Methods in Autonomous Robotic Missions,” Sensors, vol. 24, no. 4, Art. no. 4, Jan. 2024, doi: 10.3390/s24041330

    work page doi:10.3390/s24041330 2024

  28. [28]

    Multidomain EEG representation learning with orthogonal mapping and attention-based fusion for cognitive load classification,

    P. Angkan, A. Jalali, P. Hungler, and A. Etemad, “Multidomain EEG representation learning with orthogonal mapping and attention-based fusion for cognitive load classification,” IEEE Transactions on Human -Machine Systems, vol. 56, no. 2, pp. 343–355, Apr. 2026, doi: 10.1109/THMS.2026.3651706

    work page doi:10.1109/thms.2026.3651706 2026

  29. [29]

    EEG conformer: Convolutional transformer for EEG decoding and visualization,

    Y. Song, Q. Zheng, B. Liu, and X. Gao, “EEG conformer: Convolutional transformer for EEG decoding and visualization,” IEEE Trans Neural Syst Rehabil Eng , vol. 31, pp. 710 –719, 2023, doi: 10.1109/TNSRE.2022.3230250

    work page doi:10.1109/tnsre.2022.3230250 2023

  30. [30]

    Towards Unified Multi-task EEG Analysis with Low-Rank Adaptation

    S. Dai, K. Chen, H. Xiao, S. Yu, and Q. Ye, “Towards unified multi-task EEG analysis with low -rank adaptation,” arXiv preprint arXiv:2604.25131 , 2026, doi: 10.48550/arXiv.2604.25131

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.25131 2026

  31. [31]

    Progressive layered extraction (PLE): A novel multi-task learning (MTL) model for personalized recommendations,

    H. Tang, J. Liu, M. Zhao, and X. Gong, “Progressive layered extraction (PLE): A novel multi-task learning (MTL) model for personalized recommendations,” in Proceedings of the 14th ACM Conference on Recommender Systems , in RecSys ’20. New York, NY, USA: Association for Computing Machinery, Sep. 2020, pp. 269–278. doi: 10.1145/3383313.3412236

    work page doi:10.1145/3383313.3412236 2020

  32. [32]

    Modeling task relationships in multi-task learning with multi- gate mixture -of-experts,

    J. Ma, Z. Zhao, X. Yi, J. Chen, L. Hong, and E. H. Chi, “Modeling task relationships in multi-task learning with multi- gate mixture -of-experts,” in Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, in KDD ’18. New York, NY, USA: Association for Computing Machinery, Jul. 2018, pp. 1930 –1939. doi: 10.1145/3...

    work page doi:10.1145/3219819.3220007 2018

  33. [33]

    Independent component analysis of electroencephalographic data,

    S. Makeig, A. J. Bell, T. -P. Jung, and T. J. Sejnowski, “Independent component analysis of electroencephalographic data,” in Proceedings of the 9th International Conference on Neural Information Processing Systems , in NIPS’95. Cambridge, MA, USA: MIT Press, Nov. 1995, pp. 145–151

    work page 1995

  34. [34]

    Speaker Recognition from Raw Waveform with SincNet

    M. Ravanelli and Y. Bengio, “Speaker recognition from raw waveform with SincNet,” Aug. 09, 2019, arXiv: arXiv:1808.00158. doi: 10.48550/arXiv.1808.00158

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1808.00158 2019

  35. [35]

    EEGNet: a compact convolutional neural network for EEG-based brain-computer interfaces,

    V. J. Lawhern, A. J. Solon, N. R. Waytowich, S. M. Gordon, C. P. Hung, and B. J. Lance, “EEGNet: A compact convolutional neural network for EEG -based brain–computer interfaces,” J. Neural Eng., vol. 15, no. 5, Art. no. 056013, Oct. 2018, doi: 10.1088/1741-2552/aace8c

    work page doi:10.1088/1741-2552/aace8c 2018

  36. [36]

    Optimal spatial filtering of single trial EEG during imagined hand movement,

    H. Ramoser, J. Muller-Gerking, and G. Pfurtscheller, “Optimal spatial filtering of single trial EEG during imagined hand movement,” IEEE Transactions on Rehabilitation Engineering, vol. 8, no. 4, pp. 441–446, Dec. 2000, doi: 10.1109/86.895946

    work page doi:10.1109/86.895946 2000

  37. [37]

    A feature extraction and classification algorithm for motor imagery EEG signals based on decision tree and CSP-SVM,

    Y. Luo, X. He, and K. Ren, “A feature extraction and classification algorithm for motor imagery EEG signals based on decision tree and CSP-SVM,” in Optics in Health Care and Biomedical Optics XI , SPIE, Oct. 2021, p. 338. doi: 10.1117/12.2601842

    work page doi:10.1117/12.2601842 2021

  38. [38]

    Filter bank common spatial pattern (FBCSP) in brain-computer interface,

    K. K. Ang, Z. Y. Chin, H. Zhang, and C. Guan, “Filter bank common spatial pattern (FBCSP) in brain-computer interface,” in 2008 IEEE International Joint Conference on Neural Networks (IEEE World Congress on Computational Intelligence), Jun. 2008, pp. 2390 –2397. doi: 10.1109/IJCNN.2008.4634130

    work page doi:10.1109/ijcnn.2008.4634130 2008

  39. [39]

    FBCNet: A Multi-view Convolutional Neural Network for Brain-Computer Interface

    R. Mane et al. , FBCNet: A multi -view convolutional neural network for brain -computer interface . 2021. doi: 10.48550/arXiv.2104.01233

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2104.01233 2021

  40. [40]

    Deep learning with convolutional neural networks for EEG de- coding and visualization,

    R. T. Schirrmeister et al., “Deep learning with convolutional neural networks for EEG decoding and visualization,” Hum Brain Mapp, vol. 38, no. 11, pp. 5391 –5420, Aug. 2017, doi: 10.1002/hbm.23730

    work page doi:10.1002/hbm.23730 2017

  41. [41]

    Vigilance described by the time-on-task effect in EEG activity during a cued go/NoGo task,

    I. Pershin et al., “Vigilance described by the time-on-task effect in EEG activity during a cued go/NoGo task,” International Journal of Psychophysiology, vol. 183, pp. 92–102, Jan. 2023, doi: 10.1016/j.ijpsycho.2022.11.015

    work page doi:10.1016/j.ijpsycho.2022.11.015 2023

  42. [42]

    Alpha-band EEG suppression as a neural marker of sustained attentional engagement to conditioned threat stimuli,

    F. Bacigalupo and S. Luck, “Alpha-band EEG suppression as a neural marker of sustained attentional engagement to conditioned threat stimuli,” Social Cognitive and Affective Neuroscience, vol. 17, pp. 1101 –1117, Apr. 2022, doi: 10.1093/scan/nsac029

    work page doi:10.1093/scan/nsac029 2022

This paper was first reviewed by glm-5.2 on July 8, 2026.