acceptodds
Under review as a conference paper at ICLR 2027

SCoRE: Learnable Reflection Equivariance for Cross-Subject EEG Decoding

Abstract

Electroencephalography (EEG) is the principal non-invasive modality for brain-computer interfaces (BCI). Generalizing across pronounced inter-individual variability has long been a crucial challenge for cross-subject decoding. An intuitive strategy for addressing this challenge is to scale up: pretrain high-capacity models on large corpora to obtain subject-invariant representations. Yet scaling up increases model complexity and inference cost, and the shared structure such models capture remains implicit in their weights. We instead propose SCoRE, a method that remains compact while making this structure explicit through reflection equivariance, an inductive bias for EEG decoding. Physiologically, its most familiar manifestation is contralateral motor control: each hemisphere controls the opposite hand, and hand movements evoke mirror-image activity across the scalp. We posit that reflection equivariance is not confined to motor control but is intrinsic to brain activity. We operationalize this property by constraining the classifier such that a spatial reflection of the signal induces a corresponding transformation of class evidence. Specifically, SCoRE trains a shared network on each input and its reflection, constructing an equivariant classifier through joint training and parameterization. This classifier separately captures class evidence that reverses under reflection and that remains invariant. Notably, SCoRE consistently outperforms compact EEG baselines by 10.37 accuracy points and surpasses state-of-the-art pretrained models at a fraction of their size. Our method not only advances cross-subject decoding but also offers neuroscientific insight. The learned input reflection and output transformation are interpretable and agree with established physiology. SCoRE, together with EEG experiments, opens a new pathway for discovering reflection-equivariant structure across tasks and brain states.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG, brain-computer interfaces, cross-subject generalization, reflection equivariance, physiological inductive bias

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_Qbj5OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The proposed method, SCoRE, combines Euclidean alignment with shared processing of original/reflected EEG, a learned involutive input reflection, and a Householder output reflection. It guarantees logit equivariance and evaluates nine datasets, including three newly collected cohorts.

Critical strengths
  • The equivariance hypothesis is distinctive and mathematically sound.
  • Strong performance with a compact model, multiple ablations, and β-lateralization analyses make it more informative than a generic architecture study.
Critical weaknesses

(1) There is insufficient physiological justification for generalized reflection equivariance. The proposed reflection-equivariant formulation is mathematically interesting and physiologically motivated for left–right motor imagery, where bilateral neural activities underpin it. However, the manuscript extends the same principle to sleep staging, depression detection, mental arithmetic, SSVEP, and aesthetic judgment without establishing that these tasks exhibit corresponding reflection-equivariant neural representations. The reported improvements therefore support the effectiveness of a reflection-constrained learning strategy, but do not necessarily establish reflection equivariance as an intrinsic property of brain activity across tasks. A comparison with randomly initialized involutive transformations and other capacity-matched two-view learning strategies would help determine whether the improvements arise from meaningful physiological structure or more general regularization effects. (2) Need clarifications on the cross-subject generalization protocol. The method incorporates Euclidean alignment using subject/session-specific covariance estimates. The manuscript should clarify whether covariance estimation uses all target-session trials, including those subsequently evaluated. If so, the reported zero-shot results are not generalizable. Strictly prospective generalization to an unseen subject should be evaluated. The authors should distinguish these settings and ideally evaluate performance under causal or limited-calibration alignment protocols. This distinction is important for assessing the practical value of SCoRE in online BCIs and other prospective EEG applications.

Questions for authors
  1. Can the authors explain what physiological reflection-equivariant relationships are expected in these tasks (beyond motor imagery) ? In particular, how should the learned output reflection be interpreted when the class labels have no natural left–right lateralization?
  2. Have the authors compared SCoRE against randomly initialized involutive transformations or other learned orthogonal transformations with comparable model capacity?
  3. Can the authors clarify whether this covariance is computed using the entire target test session, including trials subsequently used for evaluation?
  4. How does performance change when alignment is estimated using only a limited initial calibration segment or updated causally during online decoding?
  5. For the high accuracy on the Mumtaz depression dataset, can the authors confirm that all training, validation, and test partitions are strictly subject-disjoint (given the substantial subject-specific variability and potential recording confounds in clinical EEG)?
8 comments
AuthorsOpenReview ↗

Response to Reviewer Qbj5

Thank you for your constructive comments and for recognizing the strengths of our work. Note: The revised manuscript has been uploaded, with changes marked in blue. For your convenience, the new result tables from this rebuttal are also collected in Supplementary Tables.pdf (identical to those in the revision), and the relevant ones are included in the responses below.

Q1: Reflection equivariance beyond motor imagery

A1: We would first like to respectfully clarify that the reflection in SCoRE is learned rather than predefined. Both reflections (RR and QQ) are trained jointly with the backbone, without relying on a specific electrode pairing or physiological prior. The learned reflection may pair bilateral sites, other regions, or groups of regions. We use motor imagery and bilateral pairing as our example because they are commonly understood, and then extend the concept beyond them (Section 2.2). Of course, a known or hypothesized physiological pairing can be used to initialize RR, but this is optional (initialization is discussed in Q2 below).

Physiological evidence beyond motor imagery.

TaskPhysiological evidence
Sleep stagingSlow activity is frontal, alpha is occipital, and spindles are centro-parietal, and this pattern shifts with sleep state [1, 2] (fronto-occipital, Fz–Oz).
Depression detectionDepressed participants show reduced left frontal activation [3], and the dataset's authors use interhemispheric alpha asymmetry as a feature to separate patients from controls [4] (bilateral, F3–F4).
Mental arithmeticSerial subtraction raises theta over the medial prefrontal cortex and lowers alpha over parietal regions [5] (fronto-parietal, Fz–Pz).
SSVEPThe response peak shifts between occipital and parietal sites with flicker frequency [6] (occipito-parietal, Oz–Pz).
Aesthetic judgmentFaces evoke a larger response over the right than the left posterior temporal scalp [7], and frontal alpha asymmetry relates to judgments of beauty [8] (bilateral, T5–T6 and F3–F4).

These physiological findings support the motivation of our method. They span bilateral pairs as well as fronto-occipital, fronto-parietal, and occipito-parietal pairs. Within a single task, what the model learns is likely a combination of several such relationships. We show the gain of the mechanism through performance (Table 1) and its effectiveness through ablations (Table 2). Post-hoc marker analyses further show that the model draws on physiological information (Sections 5.3 and 5.4).

To be more specific, we formalize these findings as reflection equivariance, where exchanging the paired regions changes the class evidence accordingly. The mechanism combines a built-in structure with joint learning (Section 3.2). When a task has this reflection-equivariant property, the model can capture it through this structure and gain from it.

Output reflection without left–right labels. With the clarification above, the interpretation becomes clear. QQ is not tied to left–right labels. It reverses the class contrast that the learned input reflection affects and leaves the other class evidence unchanged (Section 2.3). What matters is therefore not a predefined input–output pairing, but whether the task has a reflection-equivariant property that matches this structure. If it does, the model can in principle still learn it without left–right lateralization. If this property is weak, the gain diminishes, and performance stays comparable to strong non-pretrained baselines, as on TUAB (Appendix C.7).

Q2: Random involutions and other learned transformations

A2: As clarified in A1, the gain of SCoRE comes from its structural design, not from a physiological preset. Changing the initialization of RR (mentioned in W1) therefore does not remove this structure. In fact, our released code already offers several initializations of the input reflection. This comparison helps make the study more solid. Following your suggestion, we ran it on five datasets (Revisions: lines 1776–1781, Appendix C.12; lines 1799–1806, Table 32).

Initialization of RRBCI-IV-2aPhysioNet-MIHMCMATMumtaz
Predefined pairing (bilateral)68.49 ± 1.5663.95 ± 0.7779.79 ± 0.3078.47 ± 2.6297.44 ± 0.56
Perturbed predefined pairing67.64 ± 0.9363.57 ± 0.9079.24 ± 0.9577.88 ± 1.2897.22 ± 1.26
Random electrode pairing67.91 ± 1.7363.79 ± 0.5179.25 ± 0.9177.53 ± 3.0496.36 ± 0.87
Random involution67.90 ± 0.9364.65 ± 0.3078.71 ± 1.9778.85 ± 3.7196.11 ± 0.84

The effect of initialization exists but is minor, with differences within 1.33 points. A well-matched initialization can give slightly higher or more stable results, as the bilateral pairing does on BCI-IV-2a and Mumtaz. On PhysioNet-MI and MAT, the random involution is the best. Overall, no initialization is consistently better, and the model is robust to this configuration choice.

Other learned transformations. As you suggested, further ablations on other learned transformations and two-view strategies help clarify the source of the gain. The table below compares SCoRE with an orthogonal transformation and two-view regularization on the same five datasets (Revisions: lines 1809–1825, Appendix C.13).

MethodBCI-IV-2aPhysioNet-MIHMCMATMumtazMacro
Orthogonal transformation65.73 ± 0.6560.13 ± 0.6778.36 ± 0.2468.61 ± 3.6196.98 ± 0.5773.96
Two-view regularization66.31 ± 0.8761.69 ± 0.5675.82 ± 0.9075.76 ± 3.9095.49 ± 0.7075.01
SCoRE68.49 ± 1.5663.95 ± 0.7779.79 ± 0.3078.47 ± 2.6297.44 ± 0.5677.63

The orthogonal transformation breaks the reflection-equivariant structure. It keeps the same backbone and a comparable size, yet performance drops (macro 73.96 vs. 77.63), which indicates that the gain comes from reflection equivariance. Two-view regularization does improve on average over the backbone without the reflection mechanism (macro 75.01 vs. 74.03, Table 2). However, this improvement is a general effect of regularization, and it is much smaller than that of SCoRE over the same backbone (0.98 vs. 3.60 points). This gap indicates that regularization alone does not account for the gain of SCoRE.

Further discussion. One point beyond the scope of this study may be of interest. Initialization can be used to inject known knowledge, such as bilateral pairing. The shift of the reflection from this starting point, or the reflection learned in the end, can then be analyzed to explore physiological structure in the data. We see potential here for deep learning to offer insights to neuroscience. As this topic leans toward neuroscience and is far from the technical focus of this paper, we will not include it in the manuscript and only share it here as an insight. Thank you for your valuable suggestions, which pushed us to think more deeply and also improved this study.

Q3: Covariance computation

A3: We first clarify how the covariance is computed. In Table 1, it is computed per subject or per recording session, without labels. The same rule applies to test subjects. The key point is that the statistics used by EA are fully separated between training and test data. Training data use only training statistics, and test data use only test statistics, so there is no leakage between them (Appendix B.10 and the released code).

Generalization and causality. We understand your point (mentioned in W2) and appreciate this valuable comment. To avoid confusion and better address your concern, we would like to respectfully distinguish generalization from causality.

Generalization concerns unseen subjects. Zero-shot means using no label of the new subject for training. EA, however, is a statistical operation that involves no training. Using EA therefore keeps the evaluation zero-shot.

As for the evaluation itself, it is offline and retrospective. This follows the widely used practice in REVE [9], BIOT [10], CBraMod [11], and EEGNet [12]. It is also reasonable for these data. Several of the datasets we use have already been processed with whole-recording statistics, such as ICA-based artifact removal in the MAT release [13]. Such data are retrospective and offline by nature. Beyond BCI, tasks such as sleep staging, depression detection, and emotion recognition are also mainly analyzed offline, with little need for streaming.

Causality concerns streaming and deployment, where each prediction can only use the data received before it. A causal evaluation does make the study more solid, but it is a separate matter from generalization. We give the detailed analysis in A4.

Overall, EA uses the whole recording for its statistics in evaluation, yet the implementation involves no leakage. It is simply an offline, retrospective analysis. Other models also use such statistics and are evaluated in the same way (for example, EEG-SimpleConv computes its normalization statistics from each test session [14]). Our claim therefore remains valid, and the comparison follows the widely used protocol. To avoid confusion, we have clarified this offline setting in the revised manuscript (Revisions: lines 1127–1130, Appendix B.10), and we thank you for the helpful comment.

Reviewer_AkN9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

Motivated by physiological evidence like contralateral motor control, the paper introduces SCoRE a new model based on reflection-equivariance that performs competitively compared to the state-of-the-art EEG foundation model while benefiting from being more compact and more efficient. In addition to its model, the paper released 3 new datasets across 84 subjects. SCoRE is evaluated on the 3 new datasets introduced and 6 datasets matching the REVE evaluation protocol spanning three settings (cross-subject, personalization, and subject-dependent, and). It also provides extensive ablation study, qualitive results and computational efficiency.

Critical strengths
  • The method is well motivated, simple, and efficient while matching or outperforming recent EEG foundation models like REVE. The qualitative analysis of SCoRE explainability mechanisms matches known physiological observations.
  • The paper includes the release of 3 new datasets.
Critical weaknesses
  • The paper would benefit from additional quantitative results. For instance, 3 datasets from REVE evaluation are not reported.
Questions for authors

I would lean for a weak accept because the paper's contribution is significant by proposing an efficient and competitive model and 3 additional open-sourced datasets. The current blocker from a higher score is the evaluation that would require SCoRE to be tested on additional datasets.

Questions ordered by importance:

  • Q1: From the 10 datasets tested on REVE [1] 3 datasets (TUEV, FACED, and BCI-2020-3) are not reported in this study. Could you explain the reason they were not included? If SCoRE can be used for those datasets, could you report them using the same evaluation protocol as REVE?
  • Q2: Since the model size is a key property of the model, did you try different model sizes? Did you see saturation at bigger sizes or degradation at smaller ones?
  • Q3: To claim state-of-the-art performance for SCoRE, the method would need to compare across more datasets. Recent EEG foundation model benchmarks such as NeuralBench [2] and EEG-FM-Bench [3] would enable SCoRE to be tested on more datasets while avoiding preprocessing data, running baselines, and designing evaluation protocols.
  • Q4: In Figure 5, could you explain why you reported CPU inference latency compared to GPU inference latency or MACs (reported in Table 19)? Is inference critical in all the applications and datasets reported? In addition, could you report GPU training time? Providing this number could help users to be informed and measure the computation budget needed for training SCoRE and compare it to finetuning a pretrained EEG foundation model.

[1] El Ouahidi, Yassine, et al. "REVE: A foundation model for EEG-adapting to any setup with large-scale pretraining on 25,000 subjects." Advances in Neural Information Processing Systems 38 (2026): 22541-22577.

[2] Banville, Hubert, et al. "NeuralBench: A Unifying Framework to Benchmark NeuroAI Models." arXiv preprint arXiv:2605.08495 (2026).

[3] Xiong, Wei, et al. "EEG-FM-Bench: A Comprehensive Benchmark for the Systematic Evaluation and Diagnostic Analyses of EEG Foundation Models." arXiv preprint arXiv:2508.17742 (2025).

10 comments
AuthorsOpenReview ↗

Response to Reviewer AkN9

Thank you for your constructive comments and for recognizing the strengths of our work. Note: The revised manuscript has been uploaded, with changes marked in blue. For your convenience, the new result tables from this rebuttal are also collected in Supplementary Tables.pdf (identical to those in the revision), and the relevant ones are included in the responses below.

Q1: Three datasets from REVE's evaluation (TUEV, FACED, BCIC2020-3)

A1: Three of the datasets evaluated in REVE [1] were not included for two main reasons.

1. Consistency of the main table. The starting point of SCoRE is that reflection equivariance is common across individuals and can thus bring gains on unseen subjects. Table 1 therefore uses cross-subject evaluation with subject-disjoint splits. In REVE's protocol, BCIC2020-3 is split by trials within each participant. TUEV keeps a patient-disjoint test set, but its training and validation sets are split at the segment level. We left both out for this reason.

2. Priority for the newly collected datasets. We wanted to evaluate and release our three newly collected datasets, especially AesthEEG. Neuroaesthetics is a niche area. To our knowledge, AesthEEG is the first standardized dataset in this area that combines EEG, participants' aesthetic ratings, and large-scale independent ratings with questionnaires. The community already has common benchmarks for emotion recognition, such as FACED, but no resource of this kind. We therefore prioritized AesthEEG.

In short, space was limited, and there was no intention to exclude these datasets. Including them is important for a more solid study, so we now report all three below, following REVE's released protocol (Revisions: lines 1409–1489, Appendix C.8).

TUEVBalanced accuracy (%)Cohen's κ\kappaWeighted F1 (%)
EEGNet38.76 ± 1.430.3577 ± 0.015565.39 ± 1.20
EEGConformer40.74 ± 1.640.3967 ± 0.019569.83 ± 1.52
SPaRCNet41.61 ± 2.620.4233 ± 0.018170.24 ± 1.04
ContraWR43.84 ± 3.490.3912 ± 0.023768.93 ± 1.36
CNN-Transformer40.87 ± 1.610.3815 ± 0.013468.54 ± 2.93
FFCL39.79 ± 1.040.3732 ± 0.018867.83 ± 1.20
ST-Transformer39.84 ± 2.280.3765 ± 0.030668.23 ± 1.90
BIOT52.81 ± 2.250.5273 ± 0.024974.92 ± 0.82
LaBraM64.09 ± 0.650.6637 ± 0.009383.12 ± 0.52
CBraMod66.71 ± 1.070.6772 ± 0.009683.42 ± 0.64
REVE67.59 ± 2.290.6783 ± 0.019984.51 ± 1.29
SCoRE60.74 ± 0.860.5293 ± 0.016175.34 ± 0.76
FACEDBalanced accuracy (%)Cohen's κ\kappaWeighted F1 (%)
EEGNet40.90 ± 1.220.3342 ± 0.025141.24 ± 1.41
EEGConformer45.59 ± 1.250.3858 ± 0.018645.14 ± 1.07
SPaRCNet46.73 ± 1.550.3978 ± 0.028947.29 ± 1.33
ContraWR48.87 ± 0.780.4231 ± 0.015148.84 ± 0.74
CNN-Transformer46.97 ± 1.320.4017 ± 0.016847.20 ± 1.25
FFCL46.73 ± 1.580.3987 ± 0.038346.99 ± 1.45
ST-Transformer48.10 ± 0.790.4137 ± 0.013347.95 ± 0.96
BIOT51.18 ± 1.180.4476 ± 0.025451.36 ± 1.12
LaBraM52.73 ± 1.070.4698 ± 0.018852.88 ± 1.02
CBraMod55.09 ± 0.890.5041 ± 0.012256.18 ± 0.93
REVE56.46 ± 1.640.5080 ± 0.019156.59 ± 1.72
SCoRE53.21 ± 0.320.4711 ± 0.003652.95 ± 0.28
BCIC2020-3Balanced accuracy (%)Cohen's κ\kappaWeighted F1 (%)
EEGNet44.13 ± 0.960.3016 ± 0.012344.13 ± 1.02
EEGConformer45.06 ± 1.330.3133 ± 0.018344.88 ± 1.54
SPaRCNet44.26 ± 1.560.3033 ± 0.023344.20 ± 1.08
ContraWR42.57 ± 1.620.3078 ± 0.021844.07 ± 1.82
CNN-Transformer45.33 ± 0.920.3166 ± 0.011845.06 ± 1.27
FFCL46.78 ± 1.970.3301 ± 0.035946.89 ± 2.05
ST-Transformer41.26 ± 1.220.2941 ± 0.015942.47 ± 1.38
BIOT49.20 ± 0.860.3650 ± 0.017649.17 ± 0.79
LaBraM50.60 ± 1.550.3800 ± 0.024250.54 ± 2.05
CBraMod53.73 ± 1.080.4216 ± 0.016353.83 ± 0.96
REVE56.35 ± 1.230.4543 ± 0.015456.33 ± 1.24
SCoRE55.43 ± 1.080.4428 ± 0.013555.41 ± 1.10

These results are in line with our claims. SCoRE is not the best on these three datasets, but as a compact model it still outperforms all task-specific models and is comparable to or better than some of the pretrained models.

Q2: Model-size sensitivity

A2: We varied the model width from 0.125× to 4× of the default on BCI-IV-2a, spanning 5.8K to 4.66M parameters, and observed both degradation and saturation. The default (1×) is provided as a general-purpose setting that suits most cases. In this experiment, degradation appears only at the smallest size. This further indicates that the gain over the baselines comes from the reflection mechanism rather than from more parameters. Beyond that point, accuracy saturates and stays on a long plateau. Overall, SCoRE is robust to model size (Revisions: lines 1492–1511, Appendix C.9).

Width multiplierTemporal filters FFEmbedding dimension ddParametersBalanced accuracy (%)Cohen's κ\kappaWeighted F1 (%)
0.125×5125,83260.13 ± 1.110.4684 ± 0.014759.27 ± 1.23
0.25×102420,00767.74 ± 1.130.5699 ± 0.015167.31 ± 1.31
0.375×153643,24267.36 ± 1.350.5648 ± 0.018066.28 ± 1.45
0.5×204875,53767.77 ± 0.760.5703 ± 0.010167.08 ± 0.35
0.75×3072167,30768.60 ± 1.490.5814 ± 0.019967.90 ± 1.48
1× (default)4096295,31768.49 ± 1.560.5799 ± 0.020868.18 ± 1.43
2×801921,169,75767.46 ± 0.930.5661 ± 0.012467.60 ± 1.32
4×1603844,658,15766.55 ± 1.270.5541 ± 0.017065.98 ± 1.26

Q3: Evaluation on NeuralBench

A3: Following your suggestion, we evaluated SCoRE on NeuralBench-EEG-Core [2]. Some of its datasets require credentials or access approval. Fortunately, we already had access to almost all of them. The only exception is SEED-DV, used for the video task, which requires approval from its authors by email. We have submitted the request but have not yet received a reply. We therefore ran the remaining 35 of the 36 tasks, covering 32 of the 33 datasets, and summarize the results as in the benchmark. We downloaded 3.3 TB of data and ran all evaluations in parallel on eight servers. The results are below, with baseline results taken from the benchmark (Revisions: lines 1512–1651, Appendix C.10).

(a) Task-specific modelsMetricShallowDeep4EEGNetBDTCNEEG-CATCNetSCTACTNetSCoRE
AgePearson rr0.6970.6620.3690.5210.6670.6190.6670.6120.688
ArtifactMacro-F10.5280.4430.4550.4710.4760.4740.4910.4940.507
AudiovisualBA (%)24.20651.11128.88923.88929.64328.37359.84132.73877.619
Clinical eventMacro-F10.4580.3710.4530.3450.4540.4450.4260.4830.407
c-VEPBA (%)31.5636.30059.4744.98936.23756.84169.43380.32689.933
DementiaBA (%)41.24641.31544.00333.20244.22740.27046.44737.42052.108
DepressionBA (%)86.78386.54683.52177.04485.96480.58482.25183.64788.543
EmotionBA (%)23.85018.88523.15214.10128.16521.22818.65226.77231.938
ERNBA (%)75.26281.77479.67064.39769.63861.98182.69082.57683.411
ImageTop-5 (%)15.43357.33328.1337.90010.0335.08384.46744.88355.350
LRPBA (%)63.00179.45173.15055.01562.46256.49877.72576.24079.489
Mental arith.BA (%)69.70952.31548.47950.19866.99756.15167.06361.50872.619
Mental imageryBA (%)33.12423.33328.76221.84828.91423.50522.28628.11435.086
WorkloadBA (%)74.75441.87162.76458.84158.64263.67473.12263.78973.164
MMNBA (%)50.56559.69358.04451.50052.15450.92456.31859.81359.312
Motor executionBA (%)39.26546.28346.30317.64747.32921.53142.23855.30958.218
Motor imageryBA (%)59.59554.40258.57949.76759.70358.48961.87562.14165.474
N170BA (%)65.64273.60267.80458.12558.40357.10973.70771.79771.875
N2pcBA (%)60.36565.32662.96955.05254.77950.72962.55261.87565.234
N400BA (%)60.27864.06263.09058.09059.54951.70161.91064.54968.958
P300BA (%)51.98167.12965.18853.37457.13454.30265.35867.41868.543
Parkinson'sBA (%)63.51565.29063.22853.91469.37765.83471.50166.10562.651
PathologyBA (%)79.44278.10879.44678.31878.46680.56480.42880.11880.053
PsychopathologyPearson rr0.123—0.0440.0470.0580.1200.0370.0970.088
Reaction timePearson rr0.2860.1320.1080.0460.3360.1040.2730.3270.420
SchizophreniaBA (%)58.93569.42967.55452.77451.03672.85945.50467.19252.442
SeizureBA (%)84.38080.31584.96985.54184.25382.82179.93583.29488.288
SentenceTop-5 (%)6.8582.6524.1493.5073.8352.1105.9884.7906.501
SexBA (%)91.04489.41689.93785.58589.70389.41988.94489.97190.986
Sleep arousalBA (%)55.99658.43355.20556.07857.13157.35457.64656.89657.734
Sleep stageBA (%)68.29566.33167.53062.51268.48770.78169.94770.76270.743
SpeechTop-5 (%)3.2053.6474.3802.4624.3412.5642.8994.2152.797
SSVEPBA (%)19.0089.0678.9884.48410.6754.82177.16310.73487.083
TypingBA (%)3.6883.9103.5893.0893.8793.1603.8403.3344.884
WordTop-5 (%)4.8446.9125.1300.7465.2142.3055.7357.9388.789
Mean rank↓8.518.569.2613.469.2110.977.506.894.09
Tasks reportednn353435353535353535
Reviewer_jqFrOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Structured Contrast Reflection Equivariance (SCoRE), a model that aims to better generalize across subjects for EEG decoding tasks. The core idea is to employ a trainable equivariance structure that exploits the potential left-right (reflective) symmetry of EEG signals for some tasks. For example, for motor imagery decoding, a left-right mirroring of the EEG channels may correspond to a left-right mirroring of motor imagery classes. This learning input reflection is correspondingly paired with an output logit space reflection that is initialized via direction determined via PCA after an initial warm-up phase for training the classifier. Both reflections are constrained to be involutive, and further joint optimization is used to tune both according to a task objective. The paper conducts extensive experiments across 9 EEG datasets (including 3 new ones that they collected and provide) and compares against 11 baselines. Also, some mechanistic interpretation and evaluation of established EEG biomarkers was demonstrated.

Critical strengths

The paper is quite clearly written and overall the method and results are well presented. The methods seem well thought-out and convincingly justified.

Extensive experiments were conducted across 9 evaluation datasets (including three of their own) comparing their SCoRE method against 11 baselines. Statistical analysis was performed with significant advantages shown in most cases. The ablation studies seems fairly thorough, and the mechanistic interpretation section was nice. I appreciate that authors also reported parameters and latency comparisons, which indicates that the method is quite reasonable.

In addition, this study also provides three newly collected EEG datasets for hand motor imagery, visual stimulation, and aesthetic judgement face/landscape images. These are a substantial contribution in their own right.

I briefly checked the code and dataset links. I was happy to see they are offered via open, permissive license terms (MIT for the code, and CC BY 4.0 for the data). The code repository seems to be very well organized, with a clear README and auxiliary documentation, including requirements.txt and apparent configurations for the various experiments. However, I did not download and actually try to run the code. The datasets also appear to provide README files and supporting code.

Critical weaknesses

On line 190, the notation R0R_0 is introduced to to refer to the initial electrode-pairing matrix. However, earlier (in line 140), just RR is used to describe the construction of the initial electrode reflection matrix. I recommend making this consistent with R0R_0 as the initialization, while RR refers to the learnable input reflection.

Lines 204-206 briefly describes the learnable input reflection in words: "parameterize RR by conjugating R0R_0 with a learnable orthogonal matrix constructed via the Cayley transform". It would be helpful to clarify this with a mathematical expressions. I believe that this is applying the Cayley transform of the form (I - A)(I + A)^{-1} to constrain RR to involutive and initialized to R0R_0. I think that this should be elaborated upon, maybe in an appendix section, if space is limited in the main paper.

In Figure 5, right side, the green and blue colors are difficult to distinguish. Further, the red color may be difficult for colorblindness accessibility. I recommend considering some additional alternate way to distinguish. Maybe with an outline or differing texture?

The authors do provide their code and data via Anonymous GitHub and Figshare, however when I was initially reading the paper it was not immediately clear where exactly to find these, as the URLs do not seem to appear anywhere in the printed text. On lines 266-267, the text says "Code and the three new datasets are available on Anonymous GitHub and Figshare.", but it was not obvious to me that the text "Anonymous GitHub" and "Figshare" are actually clickable links. So, instead I went searching through the paper looking for the relevant URLs, and it took me a while to realize how to access them. Note that this is completely inaccessible with a printed copy of the paper, which is how I was initially reading it. I strongly recommend that the authors provide the URLs as printed text at least somewhere in the paper, just to make the URLs more accessible and obvious.

In the REAMDE of the anonymized code repo, the authors inadvertently left a link to their public github repo, as part of the git clone setup command. I did not recognize the github username, and did not look at their public repo or otherwise try to undermine their anonymity. I do not know the identities of the authors nor their affiliations. However, this does potentially compromise their anonymity.

Questions for authors

The EEG trial processing involves some matrix inversion and inverse square-root operations (in Equations 1 and 2) based on the second moment estimate. Were there any potential numerical stability issues with these operations in the experiments? Was there any mitigation employed?

How exactly is the learnable input reflection parameterized via the Cayley transform?

Perhaps for some tasks a reflective equivariance may not exist. Conceptually, how might the learnable reflections adapt in such a case?

4 comments
AuthorsOpenReview ↗

Response to Reviewer jqFr

Thank you for your constructive comments and for recognizing the strengths of our work. Note: The revised manuscript has been uploaded, with changes marked in blue. For your convenience, the new result tables from this rebuttal are also collected in Supplementary Tables.pdf in the Supplementary Material (identical to those in the revision), and the relevant ones are included in the responses below.

Q1: Numerical stability of the matrix operations in Eqs. (1)–(2)

A1: Positive-definite regularization is built into our Euclidean alignment (EA) implementation, and it keeps the inverse square root in Eq. (1) and the matrix inverse in Eq. (2) numerically stable across all benchmarks. We elaborate on this from the following four aspects.

1. Explicit positive-definite regularization. In the default EA implementation, eigenvalues λj\lambda_j of Σs\Sigma_s below 10−12λmax⁡10^{-12}\lambda_{\max} are raised to this threshold, giving λ^j=max⁡(λj,10−12λmax⁡)\widehat\lambda_j=\max(\lambda_j,10^{-12}\lambda_{\max}). This prevents reciprocal square roots of zero or negative eigenvalues and ensures that the regularized reference matrix Σ^s\widehat\Sigma_s and its inverse square root As=Σ^s−1/2A_s=\widehat\Sigma_s^{-1/2} are invertible. It also provides explicit condition-number bounds of κ2(Σ^s)≤1012\kappa_2(\widehat\Sigma_s)\leq10^{12} and κ2(As)≤106\kappa_2(A_s)\leq10^{6}.

2. Double-precision (float64) arithmetic. Reference matrices are accumulated in float64, explicitly symmetrized, and processed using symmetric eigendecomposition. Double-precision arithmetic reduces rounding errors during matrix construction, inversion, and multiplication. The flooring in Point 1 changes the reference matrix only slightly.

3. Preservation of the reflection structure. Positive-definite regularization preserves the algebraic construction of the reflection. Using the same invertible AsA_s and its inverse gives

Gs2=(AsRAs−1)2=AsR2As−1=IC.G_s^2=(A_sRA_s^{-1})^2=A_sR^2A_s^{-1}=I_C.

Thus, the involutive property holds in exact arithmetic for the regularized alignment operator and does not require exact identity whitening. Finite-precision accuracy is assessed separately in Point 4.

4. Empirical assessment. We checked all nine main benchmarks (table below). In practice, the reference matrices are already positive definite on seven of them, and flooring affects only about one eigenvalue per matrix on the other two. The numerical error is negligible, as even the largest relative residual between AsAs−1A_sA_s^{-1} and the identity across all benchmarks is only 4.00×10−54.00\times10^{-5}% (Revisions: line 232, Section 3.2; lines 1138–1162, Appendix B.11).

DatasetMean floored eigenvalues (%)Max. κ2(As)\kappa_2(A_s)Max. absolute inverse residualMax. relative inverse residual (%)
BCI-IV-2a0.002073.36×10−153.36\times10^{-15}4.45×10−134.45\times10^{-13}
PhysioNet-MI1.5610610^64.40×10−114.40\times10^{-11}4.26×10−94.26\times10^{-9}
HandMI0.003846.90×10−146.90\times10^{-14}2.24×10−122.24\times10^{-12}
ISRUC0.0018.96.66×10−166.66\times10^{-16}4.68×10−144.68\times10^{-14}
HMC0.002,1247.44×10−147.44\times10^{-14}6.89×10−126.89\times10^{-12}
MAT0.0065.61.11×10−151.11\times10^{-15}4.75×10−144.75\times10^{-14}
Mumtaz0.003731.05×10−141.05\times10^{-14}9.47×10−139.47\times10^{-13}
SSVEP-EEG0.003834.06×10−144.06\times10^{-14}3.00×10−123.00\times10^{-12}
AesthEEG3.2310610^61.08×10−61.08\times10^{-6}4.00×10−54.00\times10^{-5}

Q2: Cayley parameterization

A2: Thank you for the suggestion. As you correctly noted in W2, we construct the input reflection by conjugating R0R_0 with a Cayley-parameterized orthogonal matrix, which keeps RR involutive throughout training and starts it exactly at R0R_0.

1. Cayley construction. Given the initial reflection matrix R0∈RC×CR_0\in\mathbb{R}^{C\times C} with R02=ICR_0^2=I_C, we introduce C(C−1)/2C(C-1)/2 trainable parameters r\mathbf r and place them in the strictly upper-triangular part of a matrix TT, with all remaining entries set to zero. The construction is

S=T−T⊤,U=(IC−S)−1(IC+S),R=UR0U⊤.S=T-T^\top,\qquad U=(I_C-S)^{-1}(I_C+S),\qquad R=UR_0U^\top.

Skew-symmetry of SS guarantees that IC−SI_C-S is invertible and U⊤U=ICU^\top U=I_C. This is the Cayley transform in your form (I−A)(I+A)−1(I-A)(I+A)^{-1} with A=−SA=-S. In implementation, UU is computed by solving (IC−S)U=IC+S(I_C-S)U=I_C+S.

2. Initialization at R0R_0. We initialize r=0\mathbf r=\mathbf 0, giving S=0S=\mathbf 0, U=ICU=I_C, and therefore R=R0R=R_0 at the start of joint training.

3. Preservation of the involutive constraint. During joint training, gradients update r\mathbf r, allowing the reflection's spatial orientation to adapt while preserving

R2=UR0(U⊤U)R0U⊤=UR02U⊤=IC.R^2=UR_0(U^\top U)R_0U^\top=UR_0^2U^\top=I_C.

(Revisions: line 205, Section 3.2; lines 1912–1926, Appendix D)

Q3: Tasks without reflection equivariance

A3: When a physiological reflection relationship exists, the reflections learn to lock onto it and become concentrated and targeted. When it does not, the equivariance constraint still holds, and the model has to adapt to keep the equivariant structure from undermining useful evidence. Below, we compare TUAB, a task with a weak reflection relationship, with BCI-IV-2a, a task with a clear one.

Input side. On TUAB, RR spreads each channel's reflection over about seven channels, compared with fewer than two on BCI-IV-2a (7.13 vs. 1.56). At the same time, each channel draws about ten times more from itself and its neighbors.

Metric (before → after joint training)BCI-IV-2aTUAB
Effective channel count1.00 → 1.561.00 → 7.13
Coupling to the source channel (%)0.00 → 1.960.00 → 19.86
Coupling to the two nearest neighbors (%)0.00 → 2.470.00 → 24.67

Output side. On TUAB, the QQ-aligned reflected evidence moves closer to the original, and the two routes converge on the same decision (prediction agreement 13% → 72%).

Metric (before → after joint training)BCI-IV-2aTUAB
Normalized squared logit discrepancy0.464 → 0.7011.897 → 0.486
Centered-logit cosine similarity0.537 → 0.386−0.897 → 0.521
Prediction agreement (%)47.66 → 42.2712.98 → 71.56

Performance. Because such tasks lack reflection equivariance, the prior may add little gain, but owing to this adaptation, it does not hurt performance either. On TUAB, SCoRE is comparable in balanced accuracy to task-specific baselines such as ST-Transformer (79.64% vs. 79.66%). The same backbone without reflection equivariance is slightly lower, but the difference is not significant (79.33% vs. 79.64%).

An intuitive analogy is specular versus diffuse reflection. With reflection equivariance, the learned reflection acts like a mirror and sends each ray to one exact spot. Without it, the reflection acts more like a matte surface and scatters the light.

This is a high-level, conceptual explanation that draws on indirect observations across datasets, but we think it makes sense. Your question led us to think more deeply about how the model operates and what drives its behavior, which has given us real insight and pushed our thinking further. We have added the results and analysis to the revised manuscript. (Revisions: line 536, Section 6; lines 1402–1464, Appendix C.7).

Revisions for W1–W5. For W1, the electrode-pairing matrix is introduced as R0R_0, which initializes RR (Revisions: lines 140–141, Section 2.2). For W2, the Cayley parameterization is given explicitly (Revisions: line 205, Section 3.2; lines 1912–1926, Appendix D). For W3, the right panel uses a colorblind-friendly palette with distinct outline styles (Revisions: lines 505–520, Figure 8). For W4, the URLs are printed in full (Revisions: lines 555–558, Reproducibility Statement). For W5, we have updated the anonymized repository to correct this.

Reviewer_jqFrOpenReview ↗

Response to Author's Response

Thank you for detailed response. I am more convinced and confident about the quality of the paper. I maintain my maximal score.

Dear Reviewer jqFr,

Thank you again for your recognition of our study and for your continued strong support. Your confidence in this work is a great encouragement to us. Your careful and professional suggestions have also helped make it better, and we are sincerely grateful.

Have a wonderful day 😊

Best wishes,

All authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.