acceptodds
Under review as a conference paper at ICLR 2027

Multi-View Hybrid Quantum Learning with Class-Conditional Source Alignment for Cross-Subject EEG

Abstract

Cross-subject electroencephalography (EEG) learning combines few independent participants, heterogeneous task information, and distribution shift between people. Existing hybrid quantum EEG models compress all features before a single variational quantum circuit (VQC), which discards view-specific structure and leaves subject shift unaddressed. We introduce , a source-only architecture that preserves temporal, spatial, spectral, statistical, covariance, and task-dependent geometric views. Each view is processed by an independent shallow VQC and a lightweight residual encoder, a VQC-conditioned distribution weights both branches, and the fused quantum representation is regularized by class-conditional source-moment alignment and class-center separation. Under subject- or patient-disjoint evaluation on BNCI-P300, UCL ALS-EEG, and EEGET-ALS, obtains Macro-F1 scores of 86.50.5%, 66.80.5%, and 86.60.4%, exceeding the strongest of seven reimplemented architectures by 5.8, 2.0, and 11.0 points. Replacing each VQC by a LayerNorm–MLP matched to within 1% of the parameter count lowers Macro-F1 by 4.7, 3.1, and 3.0 points. Objective ablations identify conditional alignment as the largest individual contributor on BNCI and UCL, and the matched substitution isolates an additional encoder-specific increment within the aligned system. The circuits are analytically simulated bounded trigonometric feature maps; we claim no computational or hardware quantum advantage, and reported deviations quantify three-seed optimization variability under fixed outer folds.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Cross-subject EEG decoding, brain–computer interfaces, multi-view representation learning, variational quantum circuits, quantum attention

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_e237OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MVQAN-CORAL for cross-subject EEG classification. It combines multiple task-aware views, per-view quantum and classical residual encoders, shared attention, and class-conditional source alignment. Experiments on three binary EEG tasks report improvements over the evaluated baselines.

Critical strengths
  • Preserving complementary EEG views and addressing class-conditional subject shifts is a reasonable design. The reported results improve over the evaluated baselines under subject-disjoint evaluation.
Critical weaknesses

Insufficient Quantum Motivation

  • The motivation for introducing quantum encoding is insufficient. It is unclear why this component is needed for EEG. Although the paper includes an MLP-replacement ablation, it does not resolve this concern. A suitable ablation would compare quantum plus residual against residual alone, to establish the benefit of the quantum branch.

Inadequate Writing and Referencing

I think the author was writing in a rush:

  • The Introduction and Method sections contain too few references, with only one citation in the entire Method section.
  • The quantum component lacks sufficient background explanation, including in the appendix. Although Appendix E.1 lists the architecture, terms such as CNOT, RY/RZ gates, and Pauli-Z readout are not adequately introduced. This limits accessibility for readers outside quantum computing.

Limited Experimental Evaluation

  • The baselines are largely older methods and omit important recent developments. The paper lacks comparisons with recent EEG work, such as foundation models [a, b, c], geometric neural networks [f, g, h, i], and EEG alignment methods [d, e].
  • Evaluation is limited to three binary tasks, without larger-scale benchmarks such as TUAB/TUEV or any multiclass task.

References

[a] Large Brain Model for Learning Generic Representations with Tremendous EEG Data in BCI

[b] CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding

[c] REVE: A Foundation Model for EEG — Adapting to Any Setup with Large-Scale Pretraining on 25,000 Subjects

[d] SPDIM: Source-Free Unsupervised Conditional and Label Shift Adaptation in EEG

[e] HEEGNet: Hyperbolic Embeddings for EEG

[f] MAtt: A Manifold Attention Network for EEG Decoding

[g] Towards a General Attention Framework on Gyrovector Spaces for Matrix Manifolds

[h] Neural Networks on Symmetric Spaces of Noncompact Type

[i] Riemannian High-Order Pooling for Brain Foundation Models

Questions for authors

See weaknesses.

4 comments

Thank you for the specific suggestions. We have expanded the quantum explanation, added branch ablations and recent baselines, and clarified the study’s focus on EEG transfer with limited ALS patient data. The corrected implementation retains the same folds, seeds {0,1,2}, evaluation and selection settings.

W1. Quantum motivation and the residual-only comparison. The task-aware EEG views preserve ERP timing, band power and cross-channel structure. The VQC learns a compact trigonometric map with coupled coordinate terms before fusion; the primer derives z2=cos⁡a1cos⁡a2z_2=\cos a_1\cos a_2 in a two-qubit example. Its periodicity acts on learned angular coordinates, rather than directly representing EEG oscillation frequencies. The embeddings also determine the view weights shared with the residual branch. These representation and weighting functions are distinct from constructing views and aligning source distributions.

We test their contribution using encoder substitutions and the requested residual-only ablation. Residual-only removes every VQC dependency: its scorer and auxiliary losses use residual embeddings. The nominally matched variant adjusts the classical parameter budget.

Macro-F1 is mean ± sample SD across three seed-level fold averages (%):

ModelBNCIUCLEEGET
MVQAN-CORAL67.24 ± 0.2260.74 ± 0.3576.62 ± 4.39
Residual only62.74 ± 1.1047.28 ± 2.7283.63 ± 3.77
Residual only, nominally matched64.27 ± 1.0645.10 ± 5.7287.78 ± 5.97
Quantum only61.40 ± 0.2956.20 ± 0.3068.80 ± 1.10

The full model exceeds nominally matched residual-only by 2.97/15.64 points on BNCI/UCL; residual-only is stronger by 11.16 points on EEGET. With views and alignment retained, the VQC-containing configuration exceeds the MLP configuration by 6.14/6.54/8.82 points. These results support the evaluated hybrid integration, with task-dependent benefit.

Revision: §1, pp. 1–2; §3.2, pp. 3–4; §5, pp. 7–10; Appendix G, p. 18; Appendix M, p. 26, Table 21.

W2. Writing, references and quantum background. We have expanded the Introduction and Method citations and clarified the technical components at their use. The recommended works are cited as follows; letters follow your review:

Recommended worksLocation
[a–c] LaBraM, CBraMod, REVE§2, p. 3; Appendix M, p. 26
[d–e] SPDIM, HEEGNet§2, p. 3
[f–i] MAtt, GyroAtt, symmetric-space networks, RHOP§2, p. 3

The primer defines RY/RZ rotations, CNOT and Pauli-Z readout, with a worked two-qubit example. Appendix H includes the propositions’ proofs and terminal-phase lemma.

Revision: §§2–3, pp. 2–6; Appendix G, p. 18; Appendix H, p. 19.

W3. Recent baselines. We added the following evaluations. Values are mean Macro-F1 (%) across three seeds on one selected held-out unit per dataset; Table 22 includes seed SDs.

ModelBNCIUCLEEGET
LaBraM65.3849.6169.96
CBraMod70.0839.9692.89
REVE67.0147.9089.12
MAtt52.1649.1086.38
SPDNet46.0248.4878.89
TSMNet (SPDDSMBN)†54.0651.7046.41
SPDIM-style†45.4544.9436.58
EA+EEGNet†68.5248.6944.67
EA+TS-LR†51.4550.5928.88

† uses unlabeled held-out observations; our primary method is source-only. SPDIM-style denotes the evaluated adaptation variant. These single-unit results are reported separately from full-cohort averages. HEEGNet, GyroAtt, symmetric-space networks and RHOP are discussed and cited; we do not claim experimental results for them.

Revision: §4.2, p. 7; Appendix M, p. 26, Table 22.

W4. Multiclass and larger-scale evaluation. We have revised the framing to focus on source-only transfer with the limited ALS patient cohorts evaluated here. UCL addresses left/right motor-imagery decoding in eight ALS patients. EEGET provides a secondary cohort-label window task involving six ALS participants, with its home/laboratory acquisition qualification retained. BNCI2014-009 supplies a complementary P300 benchmark in ten healthy participants.

The binary endpoints follow these specific study questions. We have aligned the Abstract, Introduction, endpoint rationale and Conclusion with this scope. The present evidence concerns source-only transfer for these ALS-related tasks and the complementary healthy-participant benchmark. Multiclass or TUAB/TUEV evaluation would examine generalization beyond these endpoints.

Revision: Abstract and §1, pp. 1–2; §4.1, pp. 6–7; §§6–7, p. 10; Appendix S, p. 28.

Thank you for the response. My main concern remains unresolved: the motivation for introducing the quantum component is still unclear. From the current experiments, I don't find enough practical advantages over EEG foundation models or geometric approaches. I therefore maintain my original score.

Part 1/2: VQC motivation and ablation evidence

Thank you for clarifying your concern. We use a shallow VQC to learn coupled interactions within complementary EEG views through a compact, constrained map when few source patients are available. Its embeddings also determine view weights shared with the residual branch. Encoder, periodic, alignment and branch controls assess this architecture's task-dependent benefits.

Motivation for limited-patient transfer. Many trials still come from few independent patients: UCL uses six training, one validation and one test patient per outer fold. Temporal evolution, channel–band power and covariance remain separate views, with interactions learned before fusion.

Each source-standardized view is projected to

=\pi\tanh\!\left(g_m(\widetilde{\mathbf x}_i^{(m)})\right) \in[-\pi,\pi]^{2Q}.$$ For $Q=6$, twelve angles feed RY/RZ encoding and two trainable RY/RZ layers with ring CNOTs. The circuit parameters jointly determine the coefficients of this finite trigonometric family, including coupled coordinate functions. Appendix G derives $\langle Z_2\rangle=\cos a_1\cos a_2$. Periodic controls test alternative parameterizations. Fourier-like terms concern encoded coordinates; EEG band power is computed during view construction. Each view has an independent six-qubit, two-layer circuit. Six bounded raw Pauli-Z expectations feed learned postprocessing to produce the per-view embedding. A width-16 residual encoder retains a wider representation. Raw-readout bounds need not hold after postprocessing. *Locations: Introduction, pp. 1–2; §3.2, pp. 3–4, Eq. (2) and raw readout on p. 4; §3.5, p. 5, Proposition 1; Appendices E.1, G, H and O, pp. 16, 18, 19 and 27.* In BNCI/UCL/EEGET order (B/U/E), nominal parameter totals are 18,775/20,951/16,651 versus MLP's 18,640/20,820/16,520: less than 1% apart. Appendix I specifies model-dependent recipes and distinguishes nominal size from effective capacity. Excluding inert terminal $R_Z$ angles leaves VQC–MLP count differences of 99/89/89, not active-parameter totals. *Locations: Table 3(b), p. 8; Appendices H–I, p. 19.* **Ablations.** VQC has higher mean Macro-F1 than MLP and periodic principal-encoder controls on all three datasets, including a positive encoder difference without alignment. All ten BNCI participants, eight UCL patients and six EEGET pairs are included. Scores are Macro-F1 (%), mean±sample SD across seeds {0,1,2} after averaging fixed outer folds within each seed. Δ is full-minus-comparator pp in B/U/E order; no significance stars are assigned. |Configuration|BNCI|UCL|EEGET|Δ B/U/E| |---|---:|---:|---:|---| |**MVQAN-CORAL**|**67.24±0.22**|**60.74±0.35**|**76.62±4.39**|Reference| |MLP, alignment on|61.10±0.30|54.20±0.28|67.80±1.14|+6.14/+6.54/+8.82| |TrigRing|60.45±0.27|53.85±0.25|66.90±1.08|+6.79/+6.89/+9.72| |Fourier MLP|60.25±0.26|53.50±0.24|66.45±1.02|+6.99/+7.24/+10.17| |RFF, D=64|59.75±0.32|53.15±0.31|66.10±1.15|+7.49/+7.59/+10.52| |VQC, alignment off|60.10±0.33|55.10±0.37|67.60±1.18|+7.14/+5.64/+9.02| |MLP, alignment off|58.10±0.28|53.80±0.32|65.10±1.12|+9.14/+6.94/+11.52| |Residual only|62.74±1.10|47.28±2.72|83.63±3.77|+4.50/+13.46/−7.01| |Residual only, nominally matched|64.27±1.06|45.10±5.72|87.78±5.97|+2.97/+15.64/−11.16| |Quantum only, no residual|61.40±0.29|56.20±0.30|68.80±1.10|+5.84/+4.54/+7.82| *Locations: Tables 3–5, pp. 8–9; Appendix J, Tables 10–12, pp. 20–21; K, Table 13, p. 22; L, Table 16, p. 23; M, Table 21, p. 26.* With views and separation retained, VQC–MLP is +2.00/+1.30/+2.50 without alignment and +6.14/+6.54/+8.82 with it. Alignment adds 7.14/5.64/9.02 to VQC and 3.00/0.40/2.70 to MLP; differences of increments are 4.14/5.24/6.32. These assess encoder–regularization interaction under Appendix I's recipes. *Locations: Table 5, p. 9; Appendix L.1, p. 23; Figure 6, p. 24.* All eight held-out UCL patient means favor VQC over MLP across the same three seeds: +6.20–7.00 pp, mean +6.54. This supports improvement distributed across the evaluated ALS imagery cohort, rather than concentrated in a few patients (Table 14; Figure 5, p. 22). Removing all VQC dependencies changes view scoring and the representation receiving auxiliary losses. The hybrid exceeds both single branches on BNCI/UCL; matched residual-only is stronger on EEGET, 87.78% versus 76.62%. We retain this task dependence, without claiming the VQC essential across tasks (§5, p. 10; Appendix M, Table 21, p. 26). Together, encoder and alignment comparisons support the VQC-containing configuration within this architecture. The branch results show that its benefit is task-specific.

Part 2/2: Geometric and foundation-model comparisons

Full-cohort geometric comparisons. MVQAN-CORAL has higher mean Macro-F1 than the evaluated classical Riemannian pipelines on their corresponding full-cohort tasks. Scores below are Macro-F1 (%), mean±seed SD; parentheses give MVQAN-minus-comparator pp. — means not evaluated.

CategoryModelBNCI, 10 foldsUCL, 8 foldsEEGET, 6 folds
Proposed hybridMVQAN-CORAL67.24±0.2260.74±0.3576.62±4.39
Riemannian tangent spaceTS + SVM62.50±0.31 (+4.74)57.85±0.36 (+2.89)69.10±1.12 (+7.52)
Riemannian tangentTS + LR60.80±0.28 (+6.44)56.10±0.33 (+4.64)67.40±1.05 (+9.22)
Riemannian covariance classifierMDM56.40±0.25 (+10.84)53.20±0.30 (+7.54)62.80±0.95 (+13.82)
xDAWN+tangent spacexDAWN + TS + LR61.30±0.29 (+5.94)——
xDAWN+covariance geometryxDAWN + MDM57.80±0.26 (+9.44)——
Combined-view classifierAll-view SVM64.80±0.45 (+2.44)58.40±0.31 (+2.34)73.15±0.98 (+3.47)

Locations: Table 3(a), p. 8; Appendix J, Table 10, p. 20; Tables 11–12, p. 21.

Against TS + SVM, the strongest listed classical Riemannian pipeline, gains are 4.74/2.89/7.52 pp in BNCI/UCL/EEGET order. These compare complete pipelines; seed summaries alone do not establish paired participant-level significance.

Recent model families. LaBraM, CBraMod and REVE are EEG foundation models; MAtt/SPDNet are geometric neural networks. TSMNet combines SPD representations and domain-specific normalization. SPDIM-style and EA configurations provide additional adaptation/alignment comparisons. † uses unlabeled held-out observations; MVQAN-CORAL is source-only.

These additional methods use one selected held-out unit per dataset, averaged across three seeds, and cannot be ranked directly against MVQAN's full-cohort means above.

CategoryModelBNCIUCLEEGET
EEG foundationLaBraM65.38±1.8949.61±2.5569.96±2.60
EEG foundationREVE67.01±0.8447.90±2.2589.12±5.27
EEG foundationCBraMod70.08±3.2339.96±3.4392.89±2.98
Geometric NNMAtt52.16±1.9349.10±5.1186.38±5.05
Geometric NNSPDNet46.02±0.7048.48±4.3478.89±10.15
SPD/domain normalizationTSMNet (SPDDSMBN)†54.06±0.5951.70±0.6846.41±1.35
AdaptationSPDIM-style†45.45±0.0044.94±0.8336.58±4.11
Alignment+NNEA + EEGNet†68.52±2.8748.69±2.3244.67±6.21
Alignment+tangentEA + TS-LR†51.45±0.0050.59±0.0028.88±0.00

Scores are Macro-F1 (%), mean±sample SD across seeds for the selected unit. Foundation-model runs fine-tune public checkpoints on source participants using 200 Hz inputs and 1-s patches. SPDIM-style denotes the evaluated adaptation variant. Zero SD has Table 22's stated interpretation, not population-level certainty.

Locations: §2, p. 3; §4.2, p. 7; Appendix M, Table 22 and accompanying text, p. 26.

This panel broadens model coverage but establishes no matched full-cohort advantage over foundation models or modern geometric networks. Subtracting single-unit scores from cohort means is not a valid comparative margin; statistical testing cannot repair that mismatch. The scores also do not establish foundation-model overfitting to individual patients. Direct comparisons require matching held-out units and aggregation.

Complementary views. Geometry-only MVQAN scores 60.80/51.74/70.50% versus the full view set's 67.24/60.74/76.62%: differences of 6.44/9.00/6.12 pp. This supports combining geometry and other EEG summaries within the system; it is a view-set comparison, not an isolated quantum-effect test (Appendix N, Table 23, p. 27).

Clinical scope and practical contribution. UCL evaluates imagery transfer within eight ALS patients; BNCI is a complementary healthy-participant P300 task. EEGET predicts cohort-labeled windows, with ALS recorded at home and controls in a laboratory. This acquisition qualification applies to every model and does not identify which cues explain a particular model's score.

We substantiate a compact, source-trained hybrid representation with observed gains over the evaluated MLP/periodic configurations and full-cohort classical geometric pipelines. The Introduction and Method explain the VQC's motivation and architectural roles; the experiments show where it helps and where alternatives remain stronger. All circuits are analytically simulated. The contribution is architecture-specific, without universal encoder superiority or hardware quantum advantage.

Locations: §1, pp. 1–2; §3.2, pp. 3–4; §4.1, pp. 6–7; §5–6, pp. 7–10; Appendices Q–S, pp. 27–28. We respectfully ask that this experimental record and stated scope be considered in your final assessment.

Reviewer_w24QOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MVQAN-CORAL combines task-aware EEG views, independent variational quantum circuits (VQCs), residual encoders, attention-based fusion, and class-conditional source alignment for subject-disjoint evaluation across P300, motor imagery, and ALS-status classification tasks.

Critical strengths

The evaluation reimplements published baselines using identical data splits and includes a parameter-matched encoder substitution, objective ablations, leave-one-view-out analyses, and threshold-free metrics. Preprocessing and validation maintain subject-disjoint separation, while the authors appropriately limit their claims to an architecture-specific VQC contribution rather than quantum advantage or necessity.

Critical weaknesses

The small number of independent subjects limits the strength and generalizability of the conclusions. EEGET predictions are based on two-second EEG windows and therefore do not directly support patient-level diagnostic interpretation. More importantly, VQCs are compared only against a parameter-matched MLP, without considering classical alternatives such as random Fourier features, trigonometric networks, or kernel methods that may provide similar periodic inductive biases. Furthermore, the ablation results suggest that task-aware views and conditional alignment account for much of the performance improvement. Consequently, the findings demonstrate an incremental VQC contribution within the proposed architecture but do not establish a compelling advantage over classical EEG encoding approaches.

Questions for authors
  • What specific properties of EEG signals motivate VQC encoding beyond generic nonlinear feature transformations?
  • Can the authors compare VQCs against established classical EEG encoders and alternative methods with similar inductive biases?
  • How much performance improvement is attributable specifically to VQCs after controlling for task-aware views and conditional alignment?
1 comment

Thank you for recognizing the evaluation design. We have sharpened the EEG motivation and separated the encoder contribution from view construction and alignment. The implementation correction retains the same folds, seeds {0, 1, 2}, evaluation, and selection settings.

Q1. Why is VQC encoding suitable for EEG, and what role does it serve? P300 timing/amplitudes, motor-imagery band power and cross-channel covariance provide complementary EEG information. Our task-aware views preserve these quantities. In our architecture, per-view encoders learn nonlinear interactions among these coordinates before the views are fused.

The VQC supplies a compact, structured trigonometric map: learned angle rotations encode each view, and the CNOT network couples trigonometric functions of coordinates on different qubits. The rotation/CNOT parameterization ties the interaction coefficients, defining a structured feature family to test against MLP and explicit periodic maps. Raw Pauli-Z readouts are bounded in [-1, 1] (Proposition 1); the residual pathway retains wider classical information. The same VQC embeddings also determine the view weights used by both branches. Thus the circuit affects both the representation and which views the classifier emphasizes.

These roles are distinct from the views, which specify the input information, and alignment, which regularizes same-class representations across source participants. Neither component supplies the trainable within-view map. This explains why the encoder contribution must be tested explicitly rather than inferred from the strength of views plus alignment. We test this architectural choice against MLP, periodic and branch alternatives below.

Revision: §1, pp. 1–2; §3.2, pp. 3–4; §3.5, p. 6; Appendix G, p. 18; Appendix H, p. 19.

Q2. Classical EEG encoders and alternatives with similar inductive biases. Our complete-registry comparisons include neural EEG encoders, covariance/tangent methods, concatenated-view classifiers and periodic encoders. Selected Macro-F1 means (%) are:

ModelBNCIUCLEEGET
MVQAN-CORAL67.2460.7476.62
EEGNet58.3552.1064.20
EEG Conformer62.4555.1269.80
Tangent-space SVM62.5057.8569.10
All-view SVM64.8058.4073.15
MLP encoder61.1054.2067.80
TrigRing60.4553.8566.90
RFF, D=6459.7553.1566.10
Fourier MLP60.2553.5066.45

TrigRing uses a compressed ring-neighbour trigonometric construction; RFF uses fixed random features and a learned readout; Fourier MLP uses a learned periodic core. MVQAN-CORAL exceeds these periodic alternatives by 6.79–10.52 points across the three datasets. We compare their actual constructions without assuming identical Fourier support. The recent-model results in Appendix M are a separate single-held-out-unit evaluation, not full-cohort rankings.

Revision: §5, pp. 7–9, Tables 3–4; Appendix J, pp. 20–21; Appendix M, pp. 25–26.

Q3. Contribution with task-aware views and alignment controlled. Table 5 retains the same view definitions and separation term while separating encoder and alignment settings. Macro-F1 means (%) are:

EncoderAlignmentBNCIUCLEEGET
MLPOff58.1053.8065.10
MLPOn61.1054.2067.80
VQCOff60.1055.1067.60
VQCOn67.2460.7476.62

VQC-minus-MLP is +2.00/+1.30/+2.50 points without alignment and +6.14/+6.54/+8.82 with alignment, in BNCI/UCL/EEGET order. The encoder difference therefore remains when alignment is removed. The nominal parameter totals differ by less than 1%; training recipes and effective-capacity interpretation are specified in Appendix I.

The requested branch comparison is complementary: full-minus-nominally-matched-residual-only is +2.97/+15.64/−11.16 points. The full hybrid is stronger on BNCI/UCL, while residual-only is stronger on EEGET. Together, these results support the evaluated hybrid integration and show that its benefit is task dependent.

Revision: §5, p. 9, Table 5; Appendix I, p. 19; Appendix L, pp. 23–25; Appendix M, p. 26, Table 21.

Small cohorts and EEGET windows. The independent registry contains 10 BNCI participants, eight UCL patients and six EEGET pairs. Seed SD describes training variability. We explicitly identify EEGET as cohort-label window prediction and state its home/laboratory acquisition confound, without patient-level diagnostic interpretation.

Revision: §4.1, pp. 6–7; §6, p. 10; Appendices R–S, p. 28.

We have incorporated these explanations and comparisons into the revised paper.

Reviewer_4NdkOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes MVQAN-CORAL, a source-only framework for cross-subject EEG classification. It extracts complementary, task-aware EEG views and processes each with an independent shallow variational quantum circuit and a classical residual encoder. View weights computed from the quantum embeddings are shared across both branches, whose fused representations are concatenated for classification. Experiments on three EEG datasets report improvements over seven reimplemented architectures and a parameter-matched MLP control, supported by component ablations. The quantum circuits are analytically simulated.

Critical strengths
  1. The framework addresses relevant challenges in cross-subject EEG learning by preserving complementary task-aware views and aligning same-class representations across source subjects. The classical residual pathway also provides a coherent mechanism for retaining information compressed by the low-dimensional quantum encoders.
  2. The parameter-matched encoder replacement and objective/view ablations are useful experimental designs for assessing individual contributions. In particular, the matched replacement goes beyond comparing complete systems with substantially different architectures and parameter counts.
Critical weaknesses
  1. Some ablation results are internally inconsistent. It remains unclear whether these inconsistencies are limited to reporting or originate in the evaluation procedure. This uncertainty weakens confidence in the affected architecture comparisons.

  2. The ALS-versus-control experiment is confounded by recording environment. Patients were recorded at home, whereas healthy controls were recorded in a laboratory. Holding out participants does not remove this association, so the results cannot distinguish disease-related EEG differences from acquisition effects. This limits the evidence provided by this benchmark.

Questions for authors
  1. How consistently does the matched VQC–MLP improvement occur across held-out participants? Paired per-fold results from the existing runs would help distinguish a broadly distributed improvement from gains concentrated in a few participants, particularly on the eight-patient UCL task. A formal significance claim is not necessary.
  2. How are mini-batches constructed for the class-conditional alignment objective? With batch size 32, what are the typical numbers of subjects and samples per subject–class group contributing to Equation 6?
1 comment

Thank you for recognizing the view-preserving design and encoder comparison. We have clarified the ablation definitions, participant-wise results and alignment batches.

W1. Ablation consistency. We corrected the implementation and numerical reporting while retaining the same folds, seeds {0,1,2}, evaluation and selection settings. Repeated model entries and their differences agree across the main text and appendix. Labels now distinguish classification only, alignment off with separation retained, and simultaneous removal of alignment and the residual pathway. For the balanced UCL/EEGET test units, the reported metrics satisfy Acc = BAcc and Macro-F1 ≤ Acc.

Revision: §5, pp. 7–10, Tables 3–5; Appendix L, pp. 23–25, Tables 16–20.

W2. EEGET recording environment. We explicitly identify home acquisition for ALS and laboratory acquisition for controls. Participant holdout prevents participant overlap but does not separate cohort from recording environment. We therefore interpret EEGET as cohort-label prediction on EEG windows, without patient-level diagnostic claims or attributing its discrimination specifically to ALS physiology. UCL imagery and BNCI target detection remain separate endpoints.

Revision: §4.1, pp. 6–7, Table 2; §6, p. 10; Appendix C.3, p. 15.

Q1. Paired per-fold VQC–MLP results. We provide the numerical paired comparison for all eight UCL patients below. Each row averages the existing runs for the same held-out patient and seeds {0,1,2}. Scores are Macro-F1 (%); differences are percentage points.

Held-out patientMVQAN-CORALMLP encoderVQC − MLP
ucl_0159.2153.01+6.20
ucl_0261.1154.61+6.50
ucl_0359.2052.90+6.30
ucl_0462.0055.20+6.80
ucl_0563.0056.00+7.00
ucl_0659.5053.20+6.30
ucl_0760.4054.00+6.40
ucl_0861.5054.70+6.80
Mean60.7454.20+6.54

The patient-level seed means favor VQC in 8/8 folds, with gains of 6.20–7.00 points. The improvement is therefore distributed across the evaluated UCL patients rather than confined to a few patients. The 8/8 count concerns patient means, not individual seed runs.

For context, the complete-registry summaries are:

DatasetVQCMLP encoderMean difference
BNCI67.24 ± 0.2261.10 ± 0.30+6.14
UCL60.74 ± 0.3554.20 ± 0.28+6.54
EEGET76.62 ± 4.3967.80 ± 1.14+8.82

These are mean ± sample SD across three seed-level fold averages (%), over 10/8/6 held-out units, respectively. The numerical participant-wise comparison above concerns UCL; the BNCI/EEGET rows summarize cohort averages.

Revision: §5, pp. 7–8, Table 3(b); Appendix K, Tables 13–15 and Figure 5, pp. 22–23. Table 14 presents the numerical UCL pairs.

Q2. Mini-batch construction for Equation 6. We use subject–class-stratified sampling for filled batches of 32: select four distinct eligible training-subject pools per class and draw four distinct examples from each pool. This gives 16 examples per class, eight valid subject–class groups and four samples per group.

DatasetTraining poolDistinct subjects/batchValid groupsSamples/group
BNCI8 subjects4–884
UCL6 patients4–684
EEGET4 ALS + 4 controls884

BNCI/UCL subjects contribute both classes. The two classes’ subject sets are selected independently, so their overlap determines the distinct-subject count. EEGET subjects each carry one cohort label; its four training ALS participants and four training controls contribute to every filled batch. Validation and held-out subjects are excluded.

Each valid group receives equal outer weight, 1/8 in a filled batch. The pooled class reference uses all 16 same-class samples. The general validity rule includes only groups with at least two samples and assigns zero alignment loss to an empty valid set.

With four samples in Q = 6 dimensions, a group covariance has rank at most three. The Frobenius discrepancy is well-defined for positive-semidefinite, rank-deficient estimates and requires no matrix inversion or determinant.

Revision: §§3.3–3.4, pp. 4–5; Appendix E.3, p. 17, Equations 20–22; Appendix C.3, p. 15.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.