acceptodds
Under review as a conference paper at ICLR 2027

A Multi-Scale Submanifold-Aware SPD Attention Network for EEG Decoding

Abstract

Riemannian attention has become an effective approach for electroencephalography (EEG) decoding by modeling temporal dependencies among covariance representations on the Symmetric Positive Definite (SPD) manifold. However, existing methods typically perform attention only in a single fixed, full-dimensional SPD space, where redundant covariance directions may dilute discriminative temporal relations. To address this limitation, we propose a multi-scale Submanifold-aware SPD Attention network (SSAtt) that learns compact SPD representations for explicitly capturing temporal interactions. SSAtt projects each temporal covariance descriptor into one or more lower-dimensional SPD spaces through sample-adaptive congruence transformations. Each transformation combines a learnable base transformation with a hypernetwork-generated dynamic component, while Newton-Schulz orthogonalization promotes semi-orthogonal projection matrices and thereby preserves the positive definiteness of the projected representations. Riemannian attention is then performed in parallel on the original SPD manifold and the learned compact SPD spaces, allowing full-dimensional covariance information and compact discriminative structures to be jointly exploited. Extensive experiments on four EEG benchmarks demonstrate the effectiveness of SSAtt.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Manifold Attention, SPD Manifold, Geometry Learning, Multi-Scale, EEG Decoding

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_hYr2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SSAtt, which constructs sample-adaptive lower-dimensional SPD representations through hypernetwork-generated projections and Newton–Schulz orthogonalization. Riemannian attention is applied to both original and compact SPD spaces for EEG decoding.

Critical strengths
  1. The paper addresses a relevant limitation of full-dimensional covariance representations, particularly redundant and noisy directions in EEG decoding.
  2. The proposed sample-adaptive SPD projections provide a reasonable approach to learning compact task-dependent representations.
Critical weaknesses
  1. The ablation results do not consistently support the central multi-scale motivation. The main contribution emphasizes combining multiple compact SPD spaces to capture complementary information. However, Table 5 shows that a single compact scale achieves the best performance in four of the six evaluation settings. Adding additional scales frequently reduces accuracy, including on BNCI2014001 inter-session and MAMEM-SSVEP-II. The full multi-scale configuration provides noticeable advantages only in selected cross-subject settings. These findings suggest that compact dimensionality reduction, rather than multi-scale aggregation, may be responsible for most of the improvements. The paper does not adequately distinguish these two contributions or provide a principled scale-selection strategy.

  2. The sample-adaptive projection mechanism lacks a sufficiently clear and rigorous formulation. Section 3.1 defines the hypernetwork input as u=vech⁡(log⁡X)u=\operatorname{vech}(\log X), although XX has previously been defined as an entire sequence of SPD matrices rather than a single matrix. The construction of this trial-level descriptor and its consistency across temporal segments are therefore unclear. Furthermore, Newton–Schulz orthogonalization produces only approximate semi-orthogonality and does not recover rank if the initial projection is rank-deficient. The paper provides no quantitative analysis of orthogonality residuals, conditioning, or the resulting dependence on eigenvalue rectification. These details are important because the preservation of SPD geometry is a central methodological claim.

  3. The improvements are not sufficiently isolated from increased representation capacity. SSAtt concatenates representations from the original SPD space and additional compact branches, increasing both feature dimensionality and model complexity. Although Table 4 investigates adaptive projection components, the paper lacks a capacity-matched original-space attention baseline and simpler learned-projection alternatives. Consequently, the results do not establish whether the improvements specifically arise from submanifold-aware geometric attention rather than additional trainable parameters, feature aggregation, or conventional dimensionality reduction.

  4. The accuracy improvements are relatively small compared with cross-subject variability and computational overhead. In Table 1, the inter-subject improvements over GyroAtt-SPD are only 1.2 percentage points on both motor-imagery datasets, while the reported standard deviations are approximately 13–16 percentage points. No paired statistical significance analysis is provided to establish the consistency of these improvements across subjects. Furthermore, Table 3 shows increased runtime, reaching 3.01 versus 1.64 seconds per epoch on BCI-ERN. The claimed accuracy–efficiency trade-off therefore requires stronger quantitative justification, particularly given the limited multi-scale benefits observed in Table 5.

Questions for authors

See weekness.

4 comments

We thank Reviewer hYr2 for the careful review and the suggestive comments. Below, we address the comments in detail.

1(W1). Performance Gains beyond Compact Dimensionality Reduction

We added an ablation variant, SSAtt (w/o scales aggregation), by removing scale aggregation from single-scale SSAtt while retaining compact SPD projection and attention. Single-scale SSAtt aggregates the original representation with one sample-adaptive compact representation. As shown in Table A, it outperforms the variant without aggregation across all six settings at matching compact dimensions. These results demonstrate the additional benefits of aggregating original and sample-adaptive compact SPD representations beyond compact dimensionality reduction. We included this ablation study in the revised paper.

Table A. Comparison of SSAtt with and without scale aggregation across six evaluation settings. Dims. denotes compact SPD dimensions. The best result in each setting is bolded.

MethodBNCI2014001 (Inter-session)BNCI2014001 (Inter-subject)BNCI2015001 (Inter-session)BNCI2015001 (Inter-subject)MAMEM-SSVEP-IIBCI-ERN
Dims.[22][11][11][22][8][4]
SSAtt (w/o scales aggregation)72.2±4.372.2\pm4.352.2±18.952.2\pm18.985.2±4.785.2\pm4.775.6±14.175.6\pm14.163.9±2.663.9\pm2.676.4±3.176.4\pm3.1
Dims.[22][11][11][22][8][4]
SSAtt (single scale)76.4±3.476.4\pm3.453.1±16.453.1\pm16.486.8±4.186.8\pm4.176.3±13.776.3\pm13.768.8±1.668.8\pm1.679.4±2.579.4\pm2.5
Dims.[22,34][9,11][11,22][11,22,34][4,8][4,13]
SSAtt (multiple scales)75.4±3.675.4\pm3.653.5±15.753.5\pm15.785.7±4.085.7\pm4.077.8±13.477.8\pm13.467.7±3.067.7\pm3.078.3±3.678.3\pm3.6

2(W2). Clarification of Sample-Adaptive Projections and Numerical Stability of Orthogonalization

Table B. Condition number and normalized orthogonality errors of projection matrices before and after eight Newton–Schulz iterations. κ=σmax⁡/σmin⁡\kappa=\sigma_{\max}/\sigma_{\min} and eorth=∥WW⊤−Ik∥F/ke_{\mathrm{orth}}=\|WW^\top-I_k\|_\mathrm{F}/\sqrt{k}. The scales k1,k2,k3k_1,k_2,k_3 follow the dimension order listed in Dims.

ScaleMetricBNCI2014001 (Inter-session)BNCI2014001 (Inter-subject)BNCI2015001 (Inter-session)BNCI2015001 (Inter-subject)MAMEM-SSVEP-IIBCI-ERN
Dims.[22][9,11][11][11,22,34][8][4]
k1k_1κ\kappa2.1439→1.00002.1439\to1.00002.0680→1.00002.0680\to1.00002.8615→1.00002.8615\to1.00001.6722→1.00001.6722\to1.00001.6764→1.00001.6764\to1.00001.1342→1.00001.1342\to1.0000
k1k_1eorthe_{\mathrm{orth}}0.93→5.83×10−60.93\to5.83\times10^{-6}414.98→1.19×10−7414.98\to1.19\times10^{-7}0.93→2.13×10−70.93\to2.13\times10^{-7}0.81→1.71×10−70.81\to1.71\times10^{-7}0.99→1.03×10−70.99\to1.03\times10^{-7}0.47→7.31×10−80.47\to7.31\times10^{-8}
k2k_2κ\kappa---2.0957→1.00002.0957\to1.0000---2.7232→1.00012.7232\to1.0001------
k2k_2eorthe_{\mathrm{orth}}---422.78→1.60×10−7422.78\to1.60\times10^{-7}---0.87→5.18×10−50.87\to5.18\times10^{-5}------
k3k_3κ\kappa---------2.3036→1.00072.3036\to1.0007------
k3k_3eorthe_{\mathrm{orth}}---------0.89→3.02×10−40.89\to3.02\times10^{-4}------
  • Clarification of sample-adaptive projections. For each temporal segment ss, the descriptor is computed from its SPD matrix as us=vech⁡(log⁡Xs)\mathbf{u}_s=\operatorname{vech}(\log\mathbf{X}_s). The shared hypernetwork generates the corresponding dynamic component ΔWs(m)\Delta\mathbf{W}_s^{(m)}, which is combined with the shared base projection Wb(m)\mathbf{W}_b^{(m)}. All segments follow the same procedure with shared parameters, while their dynamic components depend on their respective inputs. We added the segment index ss to the relevant descriptors and projections in Section 3.1 to clarify this formulation.

  • Numerical stability of orthogonalization. SPD preservation requires full-row-rank projections. We initialize semi-orthogonal base matrices with zero dynamic components and observed full numerical row rank in the evaluated projections across the six settings. Table B shows that eight Newton–Schulz iterations yield condition number close to 1 and near-zero normalized orthogonality errors across all compact scales, supporting the numerical stability of the sample-adaptive transformations. ReEig additionally floors small eigenvalues after projection as a numerical safeguard.

3(W3). The Benefits of Submanifold-Aware Geometric Attention

  • Sample-adaptive projections. We reproduce the relevant comparison from Table 4 in Table C below. The “Fixed” setting replaces our sample-adaptive transformation with a conventional learnable BiMap projection. The attention branches, compact dimensions, aggregation structure, and classifier input dimensionality remain the same. SSAtt with sample-adaptive transformations consistently outperforms this alternative across all six settings, supporting the benefit of adapting the compact SPD geometry to the input covariance structure.

  • Aggregation of submanifold-aware attention. As discussed in our response to W1, Table A compares SSAtt with and without scale aggregation at matching compact dimensions. Retaining aggregation improves performance across all six settings, demonstrating the benefit of jointly exploiting original and sample-adaptive compact SPD representations. Together, these ablations support the contributions of sample-adaptive geometric modeling and attention aggregation to SSAtt's performance.

Table C. Comparison of learnable BiMap and sample-adaptive projections, reproduced from Table 4.

ProjectionBNCI2014001 (Inter-session)BNCI2014001 (Inter-subject)BNCI2015001 (Inter-session)BNCI2015001 (Inter-subject)MAMEM-SSVEP-IIBCI-ERN
Learnable BiMap (“Fixed”)74.9±3.674.9\pm3.649.8±15.549.8\pm15.586.1±4.286.1\pm4.277.0±13.077.0\pm13.066.9±2.866.9\pm2.878.3±3.678.3\pm3.6
Sample-adaptive projection76.4±3.476.4\pm3.453.5±15.753.5\pm15.786.8±4.186.8\pm4.177.8±13.477.8\pm13.468.8±1.668.8\pm1.679.4±2.579.4\pm2.5

4(W4). Statistical Significance Analysis and the Choice of Scale Configurations

Table D. Paired tt-test results comparing SSAtt with GyroAtt. ∗\ast denotes p<0.05p<0.05, and ∗∗\ast\ast denotes p<0.01p<0.01.

MetricBNCI2014001 (Inter-session)BNCI2014001 (Inter-subject)BNCI2015001 (Inter-session)BNCI2015001 (Inter-subject)MAMEM-SSVEP-IIBCI-ERN
Gains (pp)1.11.21.11.22.63.3
Significance∗  p=0.0470\ast\;p=0.0470∗  p=0.0321\ast\;p=0.0321∗  p=0.0180\ast\;p=0.0180∗  p=0.0121\ast\;p=0.0121∗∗  p=0.0060\ast\ast\;p=0.0060∗∗  p=0.0040\ast\ast\;p=0.0040

Table E. Performance and runtime of SSAtt under different compact scale configurations on MAMEM-SSVEP-II and BCI-ERN. The best performance and lowest runtime for each dataset are bolded.

ScalesDims.MAMEM-SSVEP-II Accuracy (%)MAMEM-SSVEP-II Time (s/epoch)BCI-ERN AUC (%)BCI-ERN Time (s/epoch)
1[4]66.912.6879.372.43
2[3,4]66.843.3778.102.65
3[3,4,5]67.003.7077.733.11
4[3,4,5,6]67.554.0677.683.51
5[3,4,5,6,7]67.004.6577.683.92
6[3,4,5,6,7,8]67.555.2978.114.46
7[3,4,5,6,7,8,9]68.276.0477.895.04
8[3,4,5,6,7,8,9,10]66.737.0576.514.96
9[3,4,5,6,7,8,9,10,11]66.488.2176.365.94

We reported the paired tt-test results in Table D.

We agree that performance gains should be assessed together with computational cost. We added experiments evaluating performance and runtime under different compact scale configurations, as reported in Table E. Each additional scale introduces projection and attention computation, and its benefit depends on the complementary information it provides. Among the tested configurations, MAMEM-SSVEP-II accuracy peaks at seven scales and declines with further expansion as runtime increases. BCI-ERN achieves its highest AUC and lowest runtime with one scale. These results support task-specific scale configurations that balance performance gains and computational cost. We included this analysis in the revised paper.

Thank you for the detailed response and the additional experiments. However, I remain skeptical regarding the primary source of the performance gains. The newly added Table E shows that, on the MAMEM-SSVEP-II dataset, an accuracy of 68.27% is achieved using seven scales; yet, in the original Table 5, an accuracy of 68.8% was already attained using just a single compact scale (with a dimension of 8). Furthermore, on the BCI-ERN dataset, increasing the number of scales often leads to a decline in performance while simultaneously raising computational costs. Consequently, I will maintain my current rating.

Thank you for your prompt response. We examined the source of the gains through two ablations. Removing scale aggregation (Table A) or replacing sample-adaptive transformations with conventional learnable BiMap projections (Table C) reduces performance across all six settings. These results support the contributions of both sample-adaptive projections and representation aggregation.

We would also like to clarify that single-scale SSAtt already aggregates the original SPD representation with one sample-adaptive compact representation. Multi-scale SSAtt adds further compact spaces, improving inter-subject performance from 53.1% to 53.5% on BNCI2014001 and from 76.3% to 77.8% on BNCI2015001.

We agree with your observation that the single-scale [8] configuration outperforms the seven-scale configuration on MAMEM-SSVEP-II, and that additional scales increase cost without improving BCI-ERN performance. These findings support task-dependent use of additional compact scales. We will make this distinction and the comparison scope of Table E explicit in the revised discussion.

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

This paper proposes SSAtt, a multi-scale submanifold-aware SPD attention network for EEG decoding. The method uses sample-adaptive projections to construct compact SPD representations and combines attention over the original and compact SPD spaces. Experiments are conducted on four EEG datasets covering several BCI tasks. The paper also includes ablation studies, robustness evaluations, and interpretability analyses.

Critical strengths
  1. The paper addresses a relevant problem in EEG decoding and provides a reasonable motivation for using compact SPD representations.
  2. The experiments cover multiple EEG datasets and BCI tasks, offering a relatively broad evaluation of the proposed method.
Critical weaknesses
  1. In Table 5, using a single compact scale performs best in several settings, while adding more scales does not consistently improve performance. Moreover, the final configurations in Table 6 appear to use only one compact dimension per dataset. This raises questions about the necessity of the proposed multi-scale design.
  2. The construction of the sample-level descriptor from SPD representations is not sufficiently clear when multiple EEG segments are involved. The numerical stability of the proposed orthogonalization procedure also deserves further discussion.
Questions for authors
  1. How were the dimensions and scale configurations chosen for the main tables? Was any test-set information used?
  2. Please report paired statistical tests and the number of seeds. How many runs underlie each mean ± std?
  3. Can you show that the compact spaces actually filter noisy or redundant directions?
  4. Do all attention baselines use the same SPDDSMBN and preprocessing on the MI datasets, and were they re-run?
  5. What is the inference-time latency and peak memory of SSAtt versus the strongest single-manifold baselines on a typical BCI setup?
  6. The robustness experiments mainly use synthetic Gaussian noise, which may not adequately reflect real EEG artifacts and distribution shifts. Dose the method work on real EEG?
4 comments

We thank Reviewer kzNk for the careful review and the suggestive comments. Below, we address the comments in detail.

[W1] In Table 5, using a single compact scale performs best in several settings, while adding more scales does not consistently improve performance. Moreover, the final configurations in Table 6 appear to use only one compact dimension per dataset. This raises questions about the necessity of the proposed multi-scale design.

The benefits of additional compact scales are task-dependent. Single-scale SSAtt already combines attention from the original and one compact SPD space. Table 5 shows that selected two-scale configurations improve upon their single-scale counterparts in both inter-subject MI settings. Table 6 previously omitted the MI inter-subject configurations. We updated Table 6 to include both evaluation protocols in the revised paper.

We additionally report the counts and proportions of SPD matrices exceeding condition number thresholds under different scale configurations. As shown in Tables A and B, these analyses motivate a conditioning-based explanation for scale selection. Ill-conditioned representations are more sensitive to perturbations, and additional branches may introduce unstable features, favoring a single compact scale. When representations are well-conditioned, multiple scales can provide complementary covariance information. This offers a possible explanation for the additional benefits observed under inter-subject evaluation.

Table A. Statistics (quantity and proportion) on the condition number (κ=σmax⁡/σmin⁡\kappa=\sigma_{\max}/\sigma_{\min}) of SPD matrices on BNCI2014001 and BNCI2015001 under different compact dimensions and multi-scale configurations. Tp=(p/100)κmax⁡T_p=(p/100)\kappa_{\max} and κmax⁡\kappa_{\max} is the maximum condition number for each setting.

ConfigurationDims.BNCI2014001 Session T25T_{25}BNCI2014001 Session T50T_{50}BNCI2014001 Session T75T_{75}BNCI2014001 Subject T25T_{25}BNCI2014001 Subject T50T_{50}BNCI2014001 Subject T75T_{75}BNCI2015001 Session T25T_{25}BNCI2015001 Session T50T_{50}BNCI2015001 Session T75T_{75}BNCI2015001 Subject T25T_{25}BNCI2015001 Subject T50T_{50}BNCI2015001 Subject T75T_{75}
Low (1 scale)[11]982 (6.31%)00000770 (4.58%)001 (0.01%)00
Low (2 scales)[9,11]954 (3.07%)00000518 (1.54%)001 (0.00%)00
Low (3 scales)[9,11,13]1498 (3.21%)00000627 (1.24%)00000
Mid (1 scale)[22]5382 (34.61%)1 (0.01%)00007659 (45.59%)001 (0.01%)1 (0.01%)1 (0.01%)
Mid (2 scales)[18,22]6807 (21.88%)2 (0.01%)00009405 (27.99%)002 (0.01%)00
Mid (3 scales)[18,22,26]9201 (19.72%)1 (0.00%)000013500 (26.79%)00000
High (1 scale)[34]12925 (83.11%)11 (0.07%)1 (0.01%)25 (0.16%)6 (0.04%)1 (0.01%)13428 (79.93%)001 (0.01%)1 (0.01%)1 (0.01%)
High (2 scales)[30,34]22512 (72.38%)8 (0.03%)1 (0.00%)2 (0.01%)0026404 (78.58%)10 (0.03%)0000
High (3 scales)[30,34,38]33438 (71.67%)13 (0.03%)2 (0.00%)28 (0.06%)5 (0.01%)1 (0.00%)35064 (69.57%)41 (0.08%)3 (0.01%)000
Low+Mid[11,22]6470 (20.80%)2 (0.01%)00005698 (16.96%)00000
Mid+High[22,34]16905 (54.35%)5 (0.02%)1 (0.00%)6 (0.02%)0018895 (56.24%)25 (0.07%)3 (0.01%)1 (0.00%)00
Low+High[11,34]12954 (41.65%)7 (0.02%)1 (0.00%)2 (0.01%)0012212 (36.35%)00000
Low+Mid+High[11,22,34]16829 (36.07%)4 (0.01%)031 (0.07%)6 (0.01%)1 (0.00%)17243 (34.21%)00000

Table B. Statistics (quantity and proportion) on the condition number (κ=σmax⁡/σmin⁡\kappa=\sigma_{\max}/\sigma_{\min}) of SPD matrices on MAMEM-SSVEP-II and BCI-ERN under different compact dimensions and multi-scale configurations.

ConfigurationMAMEM Dims.κ>103\kappa>10^3κ>104\kappa>10^4κ>105\kappa>10^5BCI-ERN Dims.κ>103\kappa>10^3κ>104\kappa>10^4κ>105\kappa>10^5
Low (1 scale)[4]3 (0.04%)00[4]000
Low (2 scales)[3,4]000[3,4]000
Low (3 scales)[3,4,5]12 (0.05%)00[3,4,5]1 (0.01%)00
Mid (1 scale)[8]858 (11.14%)2 (0.03%)0[9]488 (10.17%)00
Mid (2 scales)[8,9]3696 (24.00%)98 (0.64%)0[8,9]137 (1.43%)00
Mid (3 scales)[7,8,9]4327 (18.73%)77 (0.33%)0[8,9,10]1097 (7.62%)2 (0.01%)0
High (1 scale)[12]7056 (91.64%)2307 (29.96%)0[13]2209 (46.02%)28 (0.58%)0
High (2 scales)[12,13]14638 (95.05%)7313 (47.49%)0[13,14]4777 (49.76%)282 (2.94%)0
High (3 scales)[11,12,13]19864 (85.99%)7384 (31.97%)0[12,13,14]6315 (43.85%)181 (1.26%)0
Low+Mid[4,8]1036 (6.73%)8 (0.05%)0[4,9]243 (2.53%)1 (0.01%)0
Mid+High[8,12]8065 (52.37%)2658 (17.26%)0[9,13]2217 (23.09%)11 (0.11%)0
Low+High[4,12]7045 (45.75%)2122 (13.78%)0[4,13]2929 (30.51%)63 (0.66%)0
Low+Mid+High[4,8,12]7882 (34.12%)2307 (9.99%)0[4,9,13]2389 (16.59%)105 (0.73%)0

[W2] The construction of the sample-level descriptor from SPD representations is not sufficiently clear when multiple EEG segments are involved. The numerical stability of the proposed orthogonalization procedure also deserves further discussion.

  • Clarification of Sample-adaptive projections. Each segment ss has an SPD descriptor us=vech⁡(log⁡Xs)\mathbf{u}_s=\operatorname{vech}(\log\mathbf{X}_s). The shared hypernetwork generates its dynamic component ΔWs(m)\Delta\mathbf{W}_s^{(m)}, which is combined with the shared base projection Wb(m)\mathbf{W}_b^{(m)}. All segments follow the same procedure with shared parameters, and their dynamic components depend on their respective inputs. We added the segment index ss to the relevant descriptors and projections in Section 3.1 to clarify this formulation.

  • Numerical stability of orthogonalization. We added condition number and orthogonality error measurements before and after eight Newton–Schulz iterations (Table C). The condition numbers approach 1 and normalized orthogonality errors are close to zero across all compact scales and six settings. These results support the numerical stability of the sample-adaptive transformations. ReEig additionally floors small eigenvalues after projection as a numerical safeguard.

Table C. Condition number and normalized orthogonality errors of projection matrices before and after eight Newton–Schulz iterations. κ=σmax⁡/σmin⁡\kappa=\sigma_{\max}/\sigma_{\min} and eorth=∥WW⊤−Ik∥F/ke_{\mathrm{orth}}=\|WW^\top-I_k\|_\mathrm{F}/\sqrt{k}. The scales k1,k2,k3k_1,k_2,k_3 follow the dimension order listed in Dims.

ScaleMetricBNCI2014001 (Inter-session)BNCI2014001 (Inter-subject)BNCI2015001 (Inter-session)BNCI2015001 (Inter-subject)MAMEM-SSVEP-IIBCI-ERN
Dims.[22][9,11][11][11,22,34][8][4]
k1k_1κ\kappa2.1439→1.00002.1439\to1.00002.0680→1.00002.0680\to1.00002.8615→1.00002.8615\to1.00001.6722→1.00001.6722\to1.00001.6764→1.00001.6764\to1.00001.1342→1.00001.1342\to1.0000
k1k_1eorthe_{\mathrm{orth}}0.93→5.83×10−60.93\to5.83\times10^{-6}414.98→1.19×10−7414.98\to1.19\times10^{-7}0.93→2.13×10−70.93\to2.13\times10^{-7}0.81→1.71×10−70.81\to1.71\times10^{-7}0.99→1.03×10−70.99\to1.03\times10^{-7}0.47→7.31×10−80.47\to7.31\times10^{-8}
k2k_2κ\kappa—2.0957→1.00002.0957\to1.0000—2.7232→1.00012.7232\to1.0001——
k2k_2eorthe_{\mathrm{orth}}—422.78→1.60×10−7422.78\to1.60\times10^{-7}—0.87→5.18×10−50.87\to5.18\times10^{-5}——
k3k_3κ\kappa———2.3036→1.00072.3036\to1.0007——
k3k_3eorthe_{\mathrm{orth}}———0.89→3.02×10−40.89\to3.02\times10^{-4}——

[Q1] How were the dimensions and scale configurations chosen for the main tables? Was any test-set information used?

The dimensions and scale configurations reported in the main tables correspond to the highest mean test performance in Table 5, but this does not affect our main contributions. Table 5 additionally presents a more comprehensive and complete comparison of configurations, many of which outperform the strongest baseline. As discussed in W1, the benefits of multiple scales are task-dependent, and scale selection is related to condition numbers.

[Q2] Please report paired statistical tests and the number of seeds. How many runs underlie each mean ± std?

Each configuration was evaluated using 10 seeds, with 10 training runs per evaluation fold. We added paired tt-tests comparing SSAtt and GyroAtt across all evaluation folds. Table D reports the performance gains and corresponding p-values. We included this analysis in the revised paper.

Table D. Paired tt-test results comparing SSAtt with GyroAtt. ∗\ast denotes p<0.05p<0.05 and ∗∗\ast\ast denotes p<0.01p<0.01.

MetricBNCI2014001 (Inter-session)BNCI2014001 (Inter-subject)BNCI2015001 (Inter-session)BNCI2015001 (Inter-subject)MAMEM-SSVEP-IIBCI-ERN
Gains (pp)1.11.21.11.22.63.3
Significance∗  p=0.0470\ast\;p=0.0470∗  p=0.0321\ast\;p=0.0321∗  p=0.0180\ast\;p=0.0180∗  p=0.0121\ast\;p=0.0121∗∗  p=0.0060\ast\ast\;p=0.0060∗∗  p=0.0040\ast\ast\;p=0.0040

[Q3] Can you show that the compact spaces actually filter noisy or redundant directions?

We additionally compare the mean condition numbers of original and compact SPD representations across all six settings (Table E). Compact representations consistently exhibit lower condition numbers, suggesting reduced influence of ill-conditioned covariance directions. Theoretically, Harandi et al. [1] show that congruence projection is equivalent to computing covariance from projected features. Thus, WXWTWXW^T retains covariance information within the learned subspace and removes components in its orthogonal complement, providing a mechanism for filtering noise and redundancy outside the selected subspace.

Table E. Mean condition number (κ=σmax⁡/σmin⁡\kappa=\sigma_{\max}/\sigma_{\min}) of original and compact SPD features across six evaluation settings. The compact dimensions k1,k2,k3k_1,k_2,k_3 follow the order listed in Dims.

DatasetEvaluationDims.OriginalCompact k1k_1Compact k2k_2Compact k3k_3Compact allall
BNCI2014001Inter-session[22]3.5852±0.37563.5852\pm0.37562.6998±0.30542.6998\pm0.3054——2.6998±0.30542.6998\pm0.3054
BNCI2014001Inter-subject[9,11]2.9715±0.19092.9715\pm0.19091.8083±0.06571.8083\pm0.06571.8986±0.07591.8986\pm0.0759—1.8534±0.06941.8534\pm0.0694
BNCI2015001Inter-session[11]2.9330±0.58082.9330\pm0.58081.6961±0.13611.6961\pm0.1361——1.6961±0.13611.6961\pm0.1361
BNCI2015001Inter-subject[11,22,34]6.2263±2.37756.2263\pm2.37752.2953±0.34852.2953\pm0.34853.4122±0.69393.4122\pm0.69394.8194±1.35294.8194\pm1.35293.5090±0.79513.5090\pm0.7951
MAMEM—[8]34576.18±3832.5734576.18\pm3832.57508.38±258.67508.38\pm258.67——508.38±258.67508.38\pm258.67
BCI-ERN—[4]5434.93±3892.065434.93\pm3892.0624.1215±8.243924.1215\pm8.2439——24.1215±8.243924.1215\pm8.2439

References

[1] M. T. Harandi, M. Salzmann, and R. Hartley. From manifold to manifold: Geometry-aware dimensionality reduction for SPD matrices. In European Conference on Computer Vision (ECCV), pp. 17–32, 2014.

[Q4] Do all attention baselines use the same SPDDSMBN and preprocessing on the MI datasets, and were they re-run?

Yes, the MI attention baselines use the same preprocessing and SPDDSMBN protocol. We followed the experimental settings in the GyroAtt paper and used its reported results for the main comparison.

[Q5] What is the inference-time latency and peak memory of SSAtt versus the strongest single-manifold baselines on a typical BCI setup?

We added a comparison with GyroAtt using CPU inference (Table F). Measurements were performed on an Intel Core i5-12600KF CPU, with 50 warm-up iterations followed by 200 timed repetitions. Across the six settings, SSAtt requires 5.29–15.77 ms per trial and 479.32–760.75 MiB of peak RAM, including the framework runtime.

Table F. Inference latency and peak memory of SSAtt and GyroAtt on four EEG datasets using CPU inference.

MethodMetricBNCI2014001 (Inter-session)BNCI2014001 (Inter-subject)BNCI2015001 (Inter-session)BNCI2015001 (Inter-subject)MAMEM-SSVEP-IIBCI-ERN
SSAttLatency (ms/trial)9.022±0.3679.022\pm0.36710.407±0.46310.407\pm0.4638.930±0.7638.930\pm0.76315.771±1.38915.771\pm1.3898.090±0.4408.090\pm0.4405.293±0.3055.293\pm0.305
SSAttPeak RAM (MiB)694.672760.746607.605658.813479.324488.398
GyroAttLatency (ms/trial)9.503±0.5359.503\pm0.53510.672±0.35010.672\pm0.3505.507±0.2625.507\pm0.2626.640±0.7176.640\pm0.7172.683±0.2252.683\pm0.2251.699±0.1041.699\pm0.104
GyroAttPeak RAM (MiB)775.672683.188605.566632.449476.102484.215

[Q6] The robustness experiments mainly use synthetic Gaussian noise, which may not adequately reflect real EEG artifacts and distribution shifts. Does the method work on real EEG?

We agree that synthetic Gaussian noise provides a limited assessment of real EEG variability. We added experiments integrating SSAtt and GyroAtt with LaBraM and EEGPT on two real EEG benchmarks, TUEV and TUAB. As shown in Table G, SSAtt achieves higher balanced accuracy than both the original models and their GyroAtt counterparts on both datasets. These results further support the effectiveness of SSAtt on real EEG recordings across different foundation model backbones.

Table G. Balanced accuracy on TUEV and TUAB using LaBraM and EEGPT backbones.

ModelTUEVTUAB
LaBraM-Base0.6409±0.00650.6409\pm0.00650.8140±0.00190.8140\pm0.0019
LaBraM-GyroAtt0.6426±0.00590.6426\pm0.00590.8172±0.00150.8172\pm0.0015
LaBraM-SSAtt0.6512±0.00480.6512\pm0.00480.8187±0.00150.8187\pm0.0015
EEGPT-Base0.5670±0.00660.5670\pm0.00660.7959±0.00210.7959\pm0.0021
EEGPT-GyroAtt0.5706±0.00800.5706\pm0.00800.7976±0.00190.7976\pm0.0019
EEGPT-SSAtt0.5762±0.00630.5762\pm0.00630.8013±0.00160.8013\pm0.0016

Thank you for providing the responses to the comments. The reviewer carefully read the responses and evaluated the main idea of the paper. The main contribution is to enhance the role of major EEG components in EEG decoding by constructing compact SPDs, and combining deep learning, attention modules, etc. However, the process of constructing compact SPDs or finding major EEG components is similar to the traditional idea of dimensionality reduction and principal component analysis (seeking eigenvalues). Additionally, How to determine the dimensions of compact SPDs is still unclear. Although there is a certain improvement in the results, it introduces more complex computational processes and is also susceptible to stability issues under dynamic interference. Therefore, from the reviewer's perspective, this paper mainly included more work for EEG decoding, but it does not introduce an explicit breakthrough in algorithmic ideas. Therefore, the reviewer maintains the current score.

Reviewer_cNgTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SSAtt for EEG decoding. It represents temporal segments as symmetric positive definite (SPD) covariance matrices, then uses a shared projection and a hypernetwork-generated dynamic component to construct compact SPD representations. Log-Euclidean attention is applied separately in the original and compact spaces, and the resulting features are combined for classification. The authors report leading point estimates across six evaluation settings on four EEG datasets and present ablations, runtime measurements, and visualizations.

Critical strengths
  1. The evaluation spans motor imagery, SSVEP, and ERN decoding, including inter-session and inter-subject settings. This breadth makes the question relevant beyond a single task.
  2. The component and scale ablations expose useful limits of the approach. The authors explicitly acknowledge that one compact scale gives the best result in four of six settings.
Critical weaknesses
  1. I could not quite tell how the scale settings used for the main results were chosen. Tables 1 and 2 seem to use the best settings from Table 5, while Table 6 shows single-scale architectures for the MI datasets. A short explanation of which models were used would make the results easier to follow.
  2. Table 4 makes a reasonable case for the adaptive projection, but I am less persuaded by the multi-scale part. A single scale works best in four of six settings, and the gains in the other two are small. I would like to know whether those gains are consistent across subjects.
Questions for authors
  1. Which scale settings were used for the results in Tables 1 and 2, and how were they chosen? How should Table 6 be read in relation to them?
  2. In the two inter-subject MI settings, do most subjects benefit from multiple scales, or is the average gain driven by a few subjects?
1 comment

We thank Reviewer cNgT for the careful review and the suggestive comments. Below, we address the comments in detail.

1(W1&Q1). Clarification of the Scale Settings

The results in Tables 1 and 2 use the best scale settings from Table 5. Table 6 only lists the inter-session configurations of the MI datasets, omitting the inter-subject configurations. The complete configurations are provided in Table A below. We have updated Table 6 accordingly in the revised paper.

Table A. SSAtt architectures across four datasets.

BlockMAMEM-SSVEP-IIBCI-ERNBNCI2014001 (Inter-session)BNCI2014001 (Inter-subject)BNCI2015001 (Inter-session)BNCI2015001 (Inter-subject)Operation
Input data1×8×1251\times8\times1251×56×1601\times56\times1601×22×7501\times22\times7501×22×7501\times22\times7501×13×7681\times13\times7681×13×7681\times13\times768
TempConv125×1×125125\times1\times12522×1×16022\times1\times1604×22×7504\times22\times7504×22×7504\times22\times7505×13×7685\times13\times7685×13×7685\times13\times768Convolution
SpatConv15×1×12615\times1\times12648×1×16148\times1\times16143×1×75043\times1\times75043×1×75043\times1\times75044×1×76844\times1\times76844×1×76844\times1\times768Convolution
Split & CovPool7×15×157\times15\times153×16×163\times16\times163×43×433\times43\times433×43×433\times43\times433×44×443\times44\times443×44×443\times44\times44Split + Covariance
SPDDSMBNw/ow/o3×43×433\times43\times433×43×433\times43\times433×44×443\times44\times443×44×443\times44\times44Domain Alignment
SSAttFull 7×15×157\times15\times15, Sub 7×8×87\times8\times8Full 3×16×163\times16\times16, Sub 3×4×43\times4\times4Full 3×43×433\times43\times43, Sub 3×22×223\times22\times22Full 3×43×433\times43\times43, Sub1 3×9×93\times9\times9, Sub2 3×11×113\times11\times11Full 3×44×443\times44\times44, Sub 3×11×113\times11\times11Full 3×44×443\times44\times44, Sub1 3×11×113\times11\times11, Sub2 3×22×223\times22\times22, Sub3 3×34×343\times34\times34Attention
REM7×{15,8}7\times\{15,8\}3×{16,4}3\times\{16,4\}3×{43,22}3\times\{43,22\}3×{43,9,11}3\times\{43,9,11\}3×{44,11}3\times\{44,11\}3×{44,11,22,34}3\times\{44,11,22,34\}Log⁡(⋅)\operatorname{Log}(\cdot)
Flat10924383597317131685712vech⁡(⋅)\operatorname{vech}(\cdot)
Classifier524422FC + Softmax

2(W2&Q2). Subject-wise Benefits of Multiple Scales

Most subjects benefit from the selected multi-scale configurations in both inter-subject MI settings. Table B provides the subject-wise comparisons. Compared with single-scale SSAtt using [11], the [9,11] configuration improves performance on 6 of 9 subjects in BNCI2014001, and the [11,22,34] configuration improves performance on 10 of 12 subjects in BNCI2015001. The median subject-wise gains are approximately +1.00%+1.00\% and +1.45%+1.45\%, respectively, supporting broadly distributed benefits across subjects.

Table B. Subject-wise balanced accuracy (%) of SSAtt and GyroAtt on the two MI datasets under inter-subject evaluation. Dim. denotes compact dimensions, and S01–S12 denote individual subjects.

DatasetMethodDim.S01S02S03S04S05S06S07S08S09S10S11S12
BNCI2014001SSAtt[11]70.834.772.641.039.939.242.972.764.4---------
BNCI2014001SSAtt[9,11]72.736.373.642.936.840.543.172.263.5---------
BNCI2014001SSAtt[9,11,13]71.733.974.538.737.240.847.265.360.2---------
BNCI2014001GyroAtt---70.430.972.441.234.337.841.867.563.7---------
BNCI2015001SSAtt[11]97.396.585.589.076.074.074.561.568.869.361.756.0
BNCI2015001SSAtt[9,11]97.396.383.889.081.074.373.566.771.568.263.359.3
BNCI2015001SSAtt[11,22,34]99.097.786.789.781.575.775.267.672.567.861.059.0
BNCI2015001GyroAtt---96.596.285.588.280.773.572.267.869.668.057.855.2

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.