acceptodds
Under review as a conference paper at ICLR 2027

GeoLPE: Geometry-Aware Latent Prediction for EEG Foundation Modeling

11 views

Abstract

Electroencephalography (EEG) foundation models learn reusable representations across datasets and tasks, yet heterogeneous electrode montages complicate spatial transfer, while waveform reconstruction may emphasize acquisition-specific signal details. We introduce GeoLPE, a geometry-aware latent-prediction framework for EEG foundation modeling. GeoLPE combines continuous absolute spatiotemporal encoding with explicit relative inter-electrode geometry and learns target representations through structured joint-embedding prediction. We pretrain GeoLPE on 97 public EEG dataset entries and evaluate it on eight EEG datasets spanning motor imagery, event-related potentials (ERP), code-modulated visual evoked potentials (c-VEP), and cognitive workload. Among the evaluated models, GeoLPE achieves the highest cross-dataset mean balanced accuracy under all four evaluation protocols, covering cross-subject and within-subject settings with both full fine-tuning and linear probing. It reaches 69.70% under cross-subject full fine-tuning, compared with 67.54% for the strongest baseline average. The pretrained encoder is adapted to electrocorticography (ECoG) finger-trajectory prediction, achieving a mean Pearson correlation of . Ablations support the contributions of latent prediction, auxiliary objectives, and continuous geometric modeling, while representation analyses demonstrate improved transferability of pretrained representations across participants and train–test partitions. These results provide a foundation for scalable, efficient, and generalizable EEG modeling.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Electroencephalography, foundation model

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_He1qOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents GeoLPE, an EEG encoder combining continuous absolute electrode coordinates, relative geometric attention, and structured latent prediction with an EMA teacher. The pretraining objective also includes variance, waveform-reconstruction, and boundary terms. Evaluation covers eight EEG datasets under leave-one-subject-out and within-subject protocols, using both fine-tuning and frozen linear probing. A separate adaptation architecture combines the pretrained low-frequency encoder with a high-gamma branch for ECoG finger-trajectory prediction.

Critical strengths
  1. Consistent transfer improvements under two adaptation regimes. In Table 1, GeoLPE improves the cross-dataset LOSO mean over REVE from 67.54% to 69.70% with fine-tuning and from 60.25% to 62.58% with linear probing. The frozen-encoder improvement is particularly useful evidence that the representation itself carries transferable information. The within-subject results also improve on seven of eight datasets with fine-tuning and all eight with linear probing (Table 2).

  2. Substantial component analysis with appropriately qualified representation analyses. Table 4 compares waveform and latent targets, masking strategies, auxiliary losses, and geometric attention components. The held-out prototype analyses fit their transforms and class prototypes on the reference/training partition (Appendix 13.4–13.5), which makes them more informative than a visualization alone. The paper also explicitly limits its ECoG claim to later segments from seen participants and identifies published baseline numbers as non-unified comparisons.

Critical weaknesses
  1. The central geometric-transfer claim is not directly tested, and the mechanism remains only partially isolated. The eight datasets span different montages, but ordinary within-dataset evaluation does not establish robustness when the electrode layout changes between adaptation and evaluation. Table 4 shows that geometry-related components help the aggregate score; it does not show whether this benefit comes from handling layout shifts, additional capacity, or a useful inductive bias on fixed layouts. This distinction matters because montage variability is a central motivation. A focused experiment with held-out electrode subsets or layouts, comparing the full model with the existing geometry ablations, would address the claim directly. Coordinate perturbation or a shuffled-coordinate control would further distinguish geometric information from the parameterization. Per-dataset, paired uncertainty for these comparisons would also help: the aggregate ablation scores currently obscure the consistency of the effects. A matched latent-prediction comparator would strengthen novelty attribution, but a large new benchmark suite is not necessary.

  2. The ECoG experiment does not isolate the value of EEG pretraining. The reported 0.566 mean correlation is produced by a substantial adaptation system with a high-gamma branch, cross-frequency training, and three retained prediction branches (Section 3 and Appendix 11). The comparison with the published CORTEG value of 0.554 cannot establish either a reliable advantage or that the pretrained EEG encoder is responsible. The manuscript acknowledges the first limitation, but the second remains important to its cross-modality contribution. The decisive control is the same complete adaptation architecture with random initialization of the EEG encoder; a high-gamma-only control would identify how much signal comes from the added branch. Report paired participant-level differences under identical chronological splits. If these controls are unavailable, the ECoG result should be framed as feasibility of the complete adaptation system rather than evidence that EEG pretraining improves ECoG decoding.

Questions for authors
  1. Geometry clarification and targeted validation: Can the authors provide a controlled change-of-montage evaluation using the full model and the existing geometry ablations? Since the distance bias normalizes by the maximum distance in the current montage, how sensitive is the encoder to removing electrodes that change this normalization?

  2. Attribution: What is the ECoG performance of the identical three-branch adaptation system with a randomly initialized low-frequency encoder, and of a high-gamma-only variant? These results could materially change my assessment of the cross-modality contribution.

  3. Reproducibility: Can the authors provide the EEG fine-tuning/linear-probing configurations, and an executable pretraining manifest with sample counts and sampling weights? The 97 entries are explicitly not 97 independent cohorts; how are duplicate recordings and aliases across sources checked when excluding downstream data?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 3 (Part 1/4)

We thank the reviewer for the detailed assessment and suggestions for targeted controls. Below, we report controlled cross-layout evaluations and geometry ablations, coordinate-sensitivity analyses, matched ECoG controls, and additional reproducibility details.

Weakness 1 and Question 1: Controlled cross-layout evaluation

Weakness 1

The central geometric-transfer claim is not directly tested, and the mechanism remains only partially isolated. The eight datasets span different montages, but ordinary within-dataset evaluation does not establish robustness when the electrode layout changes between adaptation and evaluation. Table 4 shows that geometry-related components help the aggregate score; it does not show whether this benefit comes from handling layout shifts, additional capacity, or a useful inductive bias on fixed layouts. This distinction matters because montage variability is a central motivation. A focused experiment with held-out electrode subsets or layouts, comparing the full model with the existing geometry ablations, would address the claim directly. Coordinate perturbation or a shuffled-coordinate control would further distinguish geometric information from the parameterization. Per-dataset, paired uncertainty for these comparisons would also help: the aggregate ablation scores currently obscure the consistency of the effects. A matched latent-prediction comparator would strengthen novelty attribution, but a large new benchmark suite is not necessary.

Question 1

Geometry clarification and targeted validation: Can the authors provide a controlled change-of-montage evaluation using the full model and the existing geometry ablations? Since the distance bias normalizes by the maximum distance in the current montage, how sensitive is the encoder to removing electrodes that change this normalization?

Response to W1 and Q1

In our added cross-layout evaluation across all eight downstream datasets, GeoLPE achieves the highest BAcc on every dataset, with a macro BAcc of 61.38%, compared with 57.79% for the strongest baseline, REVE. Under the same protocol, removing spatial RoPE while retaining the geometric attention bias reduces macro BAcc by 2.64 percentage points, while removing the bias with spatial RoPE retained reduces it by 0.80 percentage points. The full model achieves macro BAcc of 64.23% under A→A and 61.38% under A→B. It achieves the highest A→B score among the evaluated geometry variants, although it does not uniformly exhibit the smallest decrease between layouts. These results support the utility of the geometric components for cross-layout performance.

We would like to clarify that “geometric-transfer” was not the central claim of our original manuscript. Our geometry-related contribution is the joint encoding of absolute spatiotemporal positions and relative inter-electrode geometry in a unified continuous coordinate system, supporting heterogeneous electrode montages during multi-dataset pretraining and downstream adaptation. Our original within-subject and LOSO evaluations across eight datasets assess pretrained representations through dataset-specific downstream adaptation, using both full fine-tuning and linear probing. Directly applying an encoder and readout adapted on one montage to another montage with all parameters fixed is a distinct evaluation setting. To address the reviewer’s question, we additionally examine this setting through controlled cross-layout evaluation and geometry ablations, and separately assess sensitivity to coordinate errors.

In a separate evaluation using the original LOSO full fine-tuning models and electrode layouts, macro BAcc decreases from 69.70% to 68.14% under test-time Gaussian coordinate perturbations with a per-axis standard deviation of 10 mm, the largest tested noise scale. This indicates limited aggregate sensitivity to the tested local coordinate errors.

For the paired A→A and A→B evaluations only, to accommodate the layout change, GeoLPE, all baselines, and the geometry ablations used mean pooling over valid electrode and temporal positions in the final encoder layer, followed by LayerNorm and a linear classifier, yielding a fixed-dimensional representation regardless of electrode count. Each encoder and readout was adapted using layout A and then evaluated on the same held-out recordings under both layout A (A→A) and layout B (A→B), with all parameters fixed during both evaluations. The same data splits and layout pairs were used for all methods and ablations. The layouts partially overlap, and layout B includes electrodes absent during adaptation, testing changes in electrode coverage while holding the dataset and task fixed. Macro BAcc is the unweighted average across the eight datasets.

AuthorsOpenReview ↗

Response to Reviewer 3 (Part 2/4)

datasetTraining layout ATest layout B
BCIC-IV-2aFC1、FCZ、FC2、C5、C3、CZ、C4、C6、CP1、CP2、P1、P2FZ、FC3、FC4、C3、C1、CZ、C2、C4、CP3、CPZ、CP4、PZ、POZ
BCIC-IV-2bC4 CZC3 CZ
BNCI2015001FC3、FC4、C5、C3、CZ、C4、C6、CPZFCZ、C3、C1、CZ、C2、C4、CP3、CP4
BNCI2014008FZ、CZ、PZ、PO7、PO8CZ、PZ、OZ、P3、P4
BNCI2014009FZ、CZ、PZ、PO7、PO8、FCZ、C3、C4、CP3、CP4CZ、PZ、OZ、P3、P4、F3、F4、C3、C4、CPZ
CastillosCVEP100FZ、F3、FC1、C3、T7、CP1、P7、O1、OZ、O2、P8、CP2、T8、C4、FC2、F4FP1、F7、FC5、C3、CP5、PZ、P3、O1、OZ、O2、P4、CP6、C4、CZ、FC6、F8、FP2
Hinss2021FP1、FZ、F3、FT9、C3、T7、CP5、CP1、PZ、P3、O1、O2、P4、CP6、CP2、FCZ、C4、T8、F4、FP2、AFZ、F5、FC3、C1、P5、PO7、PO8、P6、C2、FC4、FT10、F6FZ、F7、FC5、FC1、C3、PZ、P7、OZ、P8、C4、FT8、FC6、FC2、F8、AF7、AF3、F1、FT7、C5、TP7、CP3、P1、PO3、POZ、PO4、P2、CPZ、CP4、TP8、C6、AF8、AF4、F2
Kojima2024AAF3、AFZ、AF4、F5、F1、FZ、F2、F6、FT9、FC5、FC3、FCZ、FC4、FC6、FT10、C3、CZ、C4、TP9、CP5、CP3、CPZ、CP4、CP6、TP10、P7、P3、PZ、P4、P8、PO7、PO8、O1、O2FP1、FP2、AF7、AF8、F7、F3、FZ、F4、F8、FT7、FC1、FC2、FT8、T7、C5、C3、C1、CZ、C2、C4、C6、T8、TP7、CP1、CP2、TP8、P5、P1、PZ、P2、P6、PO3、POZ、PO4、OZ

Cross-layout evaluation

Dataset (participants)GeoLPEREVECBraModLaBraM
BCIC-IV-2a(9)42.5135.5635.2338.61
BCIC-IV-2b (9)59.6753.2656.7252.28
BNCI2015001(12)62.4458.2248.8459.23
BNCI2014008(8)61.3559.0457.8459.31
BNCI2014009(10)71.5968.6759.0768.67
CastillosCVEP100(12)76.4475.9766.4674.46
Hinss2021(15)51.3048.5244.6842.30
Kojima2024A(11)65.7163.0762.3960.15
Macro BAcc (%)61.3857.7953.9056.88

Geometry ablations under the same A→B protocol

We evaluated the existing geometry ablations under this same cross-layout setting. Fixed-frequency variants keep the scalar spatial-frequency coefficients at their deterministic initialization, fk=100−k/nf_k=100^{-k/n}, and fix the mixed-coordinate direction parameters to zero from the start of pretraining. Identity-basis variants set O=IO=I throughout training.

Spatial RoPECayley basislearnable frequenciesgeometry biasOriginal LOSO full FTWithin-layout ACross-layout A→B
w/ spatial RoPE✓✓68.8562.4860.36
w/ spatial RoPE✓✓69.2464.9861.17
w/ spatial RoPE✓✓67.8763.4660.58
w/ spatial RoPE✓67.9861.5559.92
w/ spatial RoPE✓67.2061.8959.52
w/ spatial RoPE✓67.0362.7659.13
w/ spatial RoPE✓✓✓69.7064.2361.38
w/ spatial RoPE66.5363.2959.18
w/o spatial RoPE✓66.1960.9558.74
w/o spatial RoPE64.1658.3055.32

To examine the consistency of the aggregate effects across datasets, we additionally report per-dataset A→B results for the full model and two key geometry ablations below. The w/o spatial RoPE variant retains the geometric attention bias, while the w/o geometry bias variant retains the full spatial RoPE mechanism.

datasetFull BAccw/o spatial RoPE BAccw/o geometry bias BAcc
BCIC-IV-2a42.5139.20 ± 0.1241.70 ± 0.11
BCIC-IV-2b59.6757.63 ± 0.2358.94 ± 0.06
BNCI201500162.4460.28 ± 0.1462.32 ± 0.09
BNCI201400861.3559.39 ± 0.1860.15 ± 0.14
BNCI201400971.5968.73 ± 0.1671.39 ± 0.10
CastillosCVEP10076.4472.57 ± 0.1275.21 ± 0.14
Hinss202151.3048.62 ± 0.1549.38 ± 0.06
Kojima2024A65.7163.50 ±0.0665.55 ± 0.10
Macro BAcc(%)61.3858.7460.58

Full GeoLPE achieves macro BAcc of 64.23% under A→A and 61.38% under A→B. Under A→B, the full model outperforms both ablations on all eight datasets, showing a consistent direction of the observed effects across datasets. It attains the highest A→B score among the evaluated geometry variants, although it does not uniformly exhibit the smallest decrease between layouts. The identity-basis variant performs better under A→A (64.98%) but slightly worse under A→B (61.17%), suggesting a benefit of the learned basis under the tested layout changes. These results support the utility of the geometric components for cross-layout performance, while showing that their effects on within-layout performance and sensitivity to layout changes differ.

The original-layout LOSO component ablations have been added to the Experiments section and the appendix.

AuthorsOpenReview ↗

Response to Reviewer 3 (Part 3/4)

We additionally evaluated GeoLPE’s sensitivity to coordinate errors at inference using the original LOSO full fine-tuning models and full electrode layouts. The EEG signals and all model parameters were kept fixed. We independently perturbed each electrode’s three-dimensional coordinate as pi′=pi+σϵip_i'=p_i+\sigma\epsilon_i, where ϵi∼N(0,I)\epsilon_i\sim\mathcal{N}(0,I) and σ∈2.5,5,10\sigma\in{2.5,5,10} mm denotes the standard deviation along each coordinate axis.

Test-time coordinate perturbations

Dataset (subjects)Original coordinates2.5 mm5 mm10 mm
BCIC-IV-2a (9)56.13 ± 0.9656.0555.7555.38
BCIC-IV-2b (9)76.38 ± 0.5876.1275.7274.27
BNCI2015001 (12)66.08 ± 0.6265.9765.6463.25
BNCI2014008 (8)69.35 ± 0.6569.2968.9667.65
BNCI2014009 (10)78.72 ± 0.7378.5878.0077.78
CastillosCVEP100 (12)88.19 ± 0.6188.0987.7687.19
Hinss2021 (15)53.28 ± 1.2153.0252.5551.43
Kojima2024A (11)69.44 ± 0.9269.3968.9268.13
Macro Bacc(%)69.7069.5669.1668.14

Sensitivity to distance normalization

To isolate the normalization effect, we evaluated the same layout-B recordings with the same A-adapted model, varying only the denominator of the normalized distance term. We used either the maximum pairwise electrode distance in the source layout, (DAD_A), or that in the current target layout, (DBD_B).

Normalization denominator at inferenceMacro BAcc (%)
Source-layout maximum distance (DAD_A)61.20
Current target-layout maximum distance (DBD_B)61.38

The difference is 0.18 percentage points, indicating limited aggregate sensitivity to this normalization choice under the tested layout changes. These controlled evaluations directly address the reviewer's concerns regarding montage changes, geometric sensitivity, and distance normalization, complementing the original downstream adaptation experiments.

Weakness 2 and Question 2: Contribution of EEG pretraining to ECoG decoding

Weakness 2

The ECoG experiment does not isolate the value of EEG pretraining. The reported 0.566 mean correlation is produced by a substantial adaptation system with a high-gamma branch, cross-frequency training, and three retained prediction branches (Section 3 and Appendix 11). The comparison with the published CORTEG value of 0.554 cannot establish either a reliable advantage or that the pretrained EEG encoder is responsible. The manuscript acknowledges the first limitation, but the second remains important to its cross-modality contribution. The decisive control is the same complete adaptation architecture with random initialization of the EEG encoder; a high-gamma-only control would identify how much signal comes from the added branch. Report paired participant-level differences under identical chronological splits. If these controls are unavailable, the ECoG result should be framed as feasibility of the complete adaptation system rather than evidence that EEG pretraining improves ECoG decoding.

Question 2

Attribution: What is the ECoG performance of the identical three-branch adaptation system with a randomly initialized low-frequency encoder, and of a high-gamma-only variant? These results could materially change my assessment of the cross-modality contribution.

Response to W2 and Q2

Within the same complete adaptation architecture, EEG pretraining of the low-frequency encoder increases mean Pearson rr from 0.5340 to 0.5660, an absolute gain of 0.0320, with positive paired differences for all nine participants. This matched comparison supports an additional benefit from EEG pretraining within the evaluated adaptation system. The HGA-only control achieves a mean correlation of 0.4539, below both full-system variants, showing that the complete system achieves better decoding performance than the high-gamma pathway alone. For the pretrained-versus-random comparison, both configurations retain the high-gamma branch, cross-frequency training, and all three prediction branches, and use identical downstream training settings and chronological splits. The participant-level Pearson correlations are reported below.

ParticipantPretrained full systemRandomly initialized full systemHGA-only
bp0.58170.53780.4394
cc0.6740.67260.6201
ht0.45970.41820.3188
jc0.60630.53980.4246
jp0.60170.59010.5038
mv0.71670.69480.6072
wc0.49700.45450.3928
wm0.3110.26320.2078
zt0.64590.63470.5702
Mean0.56600.5340.4539

We have added these matched ECoG controls to the Experiments section, with participant-level results and control settings in the appendix.

AuthorsOpenReview ↗

Response to Reviewer 3 (Part 4/4)

Reproducibility: Adaptation settings and pretraining provenance

Reproducibility: Can the authors provide the EEG fine-tuning/linear-probing configurations, and an executable pretraining manifest with sample counts and sampling weights? The 97 entries are explicitly not 97 independent cohorts; how are duplicate recordings and aliases across sources checked when excluding downstream data?

Response

We provide additional details below regarding the downstream adaptation settings, pretraining sampling procedure, and dataset provenance.

The configurations used for the within-subject and LOSO EEG evaluations reported in the paper are:

ConfigurationFull fine-tuningLinear probing
OptimizerAdam,(β1=0.9),(β2=0.999),(ϵ=10−8)Adam, (\beta_1=0.9), (\beta_2=0.999), (\epsilon=10^{-8})Adam,(β1=0.9),(β2=0.999),(ϵ=10−8)Adam, (\beta_1=0.9), (\beta_2=0.999), (\epsilon=10^{-8})
Learning rate5×10−45\times10^{-4}5×10−35\times10^{-3}
Weight decay0.010.001
Batch size6464
Training epochs5050
WarmupNoneExponential warmup over the first 10 epochs, applied by training step
Classification headClassification-query pooled feature together with the flattened final-layer features, followed by RMSNorm and a single linear layer, following the REVE readout design.[1]Same
Trainable parametersEncoder and classification headclassification head; encoder frozen

Pretraining manifest and sampling weights. The complete source-to-citation mapping for the 97 pretraining entries is already provided in Appendix 7.1, with the dataset provenance summary in Table 7. We will publicly release an executable pretraining manifest containing source identifiers, per-entry preprocessed segment counts, and sampling probabilities, together with the data-preparation code, to enable independent verification of the corpus composition and reproduction of the pretraining procedure. For sampling, we follow REVE[1]: we sample uniformly over the pooled set of preprocessed EEG segments.

Duplicate recordings and dataset aliases. We track dataset provenance through the original publication and data-release source associated with each entry. We check dataset aliases against this provenance mapping so that entries referring to the same original dataset or release are identified consistently even when different loader names or task-specific names are used.

References

[1] El Ouahidi Y, et al. REVE: A foundation model for EEG-adapting to any setup with large-scale pretraining on 25,000 subjects

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

In this work, authors proposed GEOLPE, a geometry-aware latent-prediction framework for EEG foundation modeling. The main contributions of this work is the combination of absolute spatiotemporal encoding with explicit relative inter-electrode geometry. Besides, they adopted a structured joint-embedding predictive architecture for EEG pretraining with focus on latent space rather than waveform space. In order to evaluate their contributions, they tested the framework across eight EEG datasets where GEOLPE achieved the best average performance among evaluated methods.

Critical strengths

The main advantages of this work are the consistence performance and sufficient experiments. First, GEOLPE achieves the highest average performance among the evaluated methods. The improvements with a frozen encoder provide particularly relevant evidence of reusable pretrained representations. Then, the ablation studies support the contribution of the spatial rotary mechanism and geometric attention bias. Furthermore, the complementary representation analyses strengthen the evaluation beyond benchmark comparisons.

Critical weaknesses

The main weakness of this work lies in the two contributions they have claimed. One contribution is the jointly encoding of absolute spatiotemporal positions and relative inter-electrode geometry. The other contribution is the specially designed JEPA architecture for EEG pretraining. However, the two contributions seem to be independent and the framework appears to be an accumulation of existing techniques.

Questions for authors
  1. Regarding the first contribution of this work where GEOLPE jointly encodes absolute spatiotemporal positions and relative inter-electrode geometry in a unified continuous coordinate system, what additional capabilities arise from the learnable spatial frequencies, Cayley-parameterized orthogonal basis, and geometric attention bias?
  2. The two mappings used to produce the positional embedding seem to be redundancy? Can you explain why the fixed Fourier mapping and the learnable coordinate mapping are both used?
  3. Please clarify the subject- and recording-level overlap between the 97 pretraining dataset entries and the eight downstream datasets? I wonder if the evaluation has directly established generalization to unseen data sources.
  4. In order to prove the generalization ability of the framework, authors should add an experiment to test the performance across the datasets. Currently, the results are tested within the same datasets, the performance tested across the datasets has not displayed yet.
7 comments
AuthorsOpenReview ↗

Response to Reviewer 2 (Part 1/3)

We thank the reviewer for the thoughtful questions and constructive feedback. Below, we clarify the connection between geometry and latent prediction, examine the positional components, document dataset provenance, and report additional cross-dataset transfer results.

Weakness 1: Connection between geometry and latent prediction

The main weakness of this work lies in the two contributions they have claimed. One contribution is the jointly encoding of absolute spatiotemporal positions and relative inter-electrode geometry. The other contribution is the specially designed JEPA architecture for EEG pretraining. However, the two contributions seem to be independent and the framework appears to be an accumulation of existing techniques.

Response to W1

GeoLPE builds on established positional encoding and latent-prediction methods. Its contribution is an EEG-specific framework in which electrode geometry structures both feature aggregation and the predictive learning task.

Specifically, the continuous coordinate system guides spatial attention, defines spatial-cluster masks and spatial predictor neighborhoods, provides positional embeddings for masked prediction queries, and determines boundary-alignment weights. The predictive objective trains these geometry-conditioned representations to recover reference-view latent representations from incomplete observations with coordinate and amplitude perturbations. Thus, geometry guides both feature aggregation and the construction of the predictive learning task. Our matched uniform-weighting control and controlled cross-layout geometry ablations provide complementary evidence for the usefulness of these EEG-specific adaptations.

We have added the matched boundary-weighting control and the original-layout LOSO geometry-component ablations to the Experiments section and the appendix of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer 2 (Part 2/3)

Question 1: Roles of the geometric components

Regarding the first contribution of this work where GEOLPE jointly encodes absolute spatiotemporal positions and relative inter-electrode geometry in a unified continuous coordinate system, what additional capabilities arise from the learnable spatial frequencies, Cayley-parameterized orthogonal basis, and geometric attention bias?

Response to Q1

Beyond providing an absolute positional reference, these components enable two complementary capabilities for cross-channel aggregation. We evaluated the corresponding pretraining ablations under LOSO full fine-tuning across eight datasets, using the full model as the reference.

  1. Feature-dependent spatial matching. Learnable spatial frequencies determine the spatial directions and scales to which Q/K matching is sensitive, while the Cayley-parameterized orthogonal basis learns which feature combinations receive each rotation. Together, they form a spatial RoPE mechanism that conditions feature matching on relative electrode geometry. Fixing the spatial-frequency parameterization and replacing the learned basis with the identity reduce macro BAcc by 0.85 and 0.46 percentage points, respectively, supporting the benefit of adapting both spatial sensitivity and the feature basis.
  2. Content-independent geometric preferences. The geometric attention bias maps electrode-pair direction and normalized distance to an additive attention score, providing an explicit spatial preference independent of the current signal content. Removing this bias reduces macro BAcc by 1.83 percentage points even when spatial RoPE is retained, indicating that these explicit geometric preferences provide additional value alongside feature-dependent rotary matching.

Fixed-frequency variants keep the scalar spatial-frequency coefficients at their deterministic initialization, fk=100−k/nf_k=100^{-k/n}, and fix the mixed-coordinate direction parameters to zero from the start of pretraining. Identity-basis variants set O=IO=I throughout training. Macro BAcc is the unweighted average across the eight datasets.

Spatial RoPECayley basislearnable frequenciesgeometry biasMacro BAcc
w/ spatial RoPE✓✓68.85
w/ spatial RoPE✓✓69.24
w/ spatial RoPE✓✓67.87
w/ spatial RoPE✓67.98
w/ spatial RoPE✓67.20
w/ spatial RoPE✓67.03
w/ spatial RoPE✓✓✓69.70
w/ spatial RoPE66.53
w/o spatial RoPE✓66.19
w/o spatial RoPE64.16

Removing spatial RoPE while retaining the geometric attention bias reduces macro BAcc from 69.70% to 66.19%, a decrease of 3.51 percentage points. Together, these results support the contributions of feature-dependent geometric matching and content-independent spatial preferences within the shared aggregation framework.

We have expanded the discussion of these components in the Experiments section and added the complete ablation results and control definitions to the appendix.

Question 2: Complementarity of the positional mappings

Response to Q2

Combining the two mappings achieves macro BAcc of 69.70% under LOSO and 63.73% under within-subject full fine-tuning, outperforming either mapping alone under both protocols.

Both mappings take coordinates as input and jointly produce the absolute spatiotemporal embedding. The fixed Fourier mapping provides an explicit multiscale geometric prior, while the learnable coordinate mapping adapts the positional representation to the EEG pretraining objective. Their combination provides a predefined multiscale structure together with a learned positional adjustment.

The table reports the mean BAcc (%) across all eight downstream datasets.

Positional encodingLOSOWithin-subject
Fourier + learnable mapping69.763.73
Fourier-only67.4261.44
Learnable-only67.9661.74

Adding the learnable mapping to Fourier-only improves macro BAcc by 2.28 and 2.29 percentage points under LOSO and within-subject evaluation, respectively, supporting the benefit of learned positional adjustment. Adding the fixed Fourier mapping to learnable-only yields gains of 1.74 and 1.99 percentage points, supporting the value of retaining the explicit multiscale prior. These comparisons support complementary contributions from both mappings in the evaluated settings.

AuthorsOpenReview ↗

Response to Reviewer 2 (Part 3/3)

Question 3: Pretraining and downstream data overlap

Please clarify the subject- and recording-level overlap between the 97 pretraining dataset entries and the eight downstream datasets? I wonder if the evaluation has directly established generalization to unseen data sources.

Response to Q3

At the dataset-entry level, there is no overlap between the eight downstream EEG datasets and the 97 pretraining entries. Appendix 7.1 documents the original publications and data-release sources for the pretraining corpus, while Appendix Table 8 lists the downstream datasets. We check dataset aliases and task-specific variants against this provenance mapping to identify entries originating from the same release. We additionally performed a code-based overlap check, comparing every trial from each participant in the eight downstream datasets against all trials in the pretraining corpus using sample-wise signal subtraction. No identical trial signals were detected. We will publicly release our implementation, pretraining manifest, and data-preparation and evaluation scripts, together with the source mapping and split definitions, to enable independent verification of the dataset composition and reproduction of the experiments. The additional leave-one-dataset-out experiments in our response to Question 4 further evaluate direct transfer to datasets absent from pretraining.

Question 4: Cross-dataset generalization

In order to prove the generalization ability of the framework, authors should add an experiment to test the performance across the datasets. Currently, the results are tested within the same datasets, the performance tested across the datasets has not displayed yet.

Response to Q4

GeoLPE achieves the highest BAcc on all five held-out targets, with a macro BAcc of 64.41%, compared with 61.08% for the strongest baseline, REVE. With the task and label space held fixed, these results support direct reuse of the source-trained encoder and classifier across datasets without target-specific fine-tuning or classifier refitting.

Our original claims concern the generalization of pretrained EEG representations under within-subject and cross-subject LOSO evaluation, supported by full fine-tuning and linear probing results across eight datasets. Direct transfer to a held-out dataset with both the source-trained encoder and classification head fixed extends beyond the original evaluation scope. To assess this additional setting, we conducted a five-dataset leave-one-dataset-out MI evaluation.

We standardized the datasets to left-hand versus right-hand motor-imagery classification, retaining only the corresponding trials in BCIC-IV-2a and Stieger2021. In each fold, we performed supervised training on four source datasets and directly evaluated the source-trained encoder and classification head on the remaining target dataset, with all parameters fixed during target evaluation. This procedure was repeated across all five targets.

Held-out target datasetGeolpeREVECBraModLaBraM
BCIC-IV-2a67.0164.9151.9362.54
BCIC-IV-2b58.9953.7552.9951.54
Dreyer202362.8556.2553.4556.13
Stieger202169.0766.6362.3350.14
SHU-MI64.1263.8859.2356.78
Macro BAcc (%)64.4161.0855.9955.43

Values are BAcc (%). Macro BAcc is the unweighted mean across the five held-out target datasets. These additional results complement the original evaluations and further support the transferability of GeoLPE’s pretrained representations.

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

This paper introduces GeoLPE, a self-supervised EEG pretraining framework designed for heterogeneous electrode montages. It combines continuous spatiotemporal positional encoding with explicit relative electrode geometry in spatial attention, and uses structured masked latent prediction with auxiliary regularization and waveform reconstruction. Evaluations on eight EEG datasets show higher average balanced accuracy than the compared baselines under cross-subject and within-subject settings, using both fine-tuning and linear probing. The paper also explores adaptation to ECoG finger-trajectory prediction.

Critical strengths
  1. The results support the practical value of the pretrained representations. Improvements under both fine-tuning and frozen-encoder linear probing suggest that GeoLPE learns features that transfer across several EEG task families. Using one shared pretrained checkpoint across heterogeneous datasets further supports its usefulness as a reusable encoder.
  2. The integration is coherent and supported by relevant ablations. Continuous electrode geometry and structured latent prediction address complementary aspects of EEG representation learning: spatial relationships across recording layouts and dependencies across partially observed signals. Within-framework ablations support the usefulness of both components, giving the proposed combination empirical value despite its reliance on established ideas.
Critical weaknesses
  1. The methodological advance is incremental, and the value of the EEG-specific modifications is only partially established. Continuous coordinate encoding, spatial rotary mechanisms, and EMA-based latent prediction have clear precedents in REVE, STRING, and EEG2Rep. The contribution therefore rests primarily on their integration and the specific adaptations. Table 4 supports the usefulness of the complete recipe, but does not fully isolate the value of these adaptations. For example, removing the boundary loss also removes all visible-token alignment, leaving unclear whether the gain comes from boundary-aware weighting or simply additional supervision.
  2. The evidence for improved transfer across electrode layouts remains indirect. The geometry ablations in Table 4 establish usefulness within the evaluated settings, but LOSO primarily tests participant transfer, while comparisons across datasets jointly change the task, signal distribution, and montage.
Questions for authors
  1. On the same held-out recordings, how does performance change when peripheral electrodes are removed, using the existing encoder and readout?
  2. Does the boundary-loss gain persist when the distance-dependent weights in Eq. (9) are replaced by their per-record mean, preserving the overall loss weight and training schedule?
6 comments
AuthorsOpenReview ↗

Response to Reviewer 1 (Part 1/4)

We thank the reviewer for the careful assessment and constructive feedback. We address the comments below with a matched boundary-weighting control, cross-dataset transfer experiments, and a same-recording electrode-removal evaluation.

Weakness 1 and Question 2: Boundary-aware weighting

Weakness 1

The methodological advance is incremental, and the value of the EEG-specific modifications is only partially established. Continuous coordinate encoding, spatial rotary mechanisms, and EMA-based latent prediction have clear precedents in REVE, STRING, and EEG2Rep. The contribution therefore rests primarily on their integration and the specific adaptations. Table 4 supports the usefulness of the complete recipe, but does not fully isolate the value of these adaptations. For example, removing the boundary loss also removes all visible-token alignment, leaving unclear whether the gain comes from boundary-aware weighting or simply additional supervision.

Question 2

Does the boundary-loss gain persist when the distance-dependent weights in Eq. (9) are replaced by their per-record mean, preserving the overall loss weight and training schedule?

Response to W1 and Q2

Boundary-aware weighting outperforms the matched uniform-weighting control on all eight datasets under both evaluation protocols, improving macro BAcc by 1.51 percentage points under LOSO and 1.69 percentage points under within-subject evaluation.

To construct the uniform-weighting control, we replace the distance-dependent weights in Eq. (9) with their per-record mean over visible tokens. The loss normalization, overall loss coefficient, and training schedule are matched between the uniform-weighting and boundary-weighting conditions. We compare three configurations during pretraining and evaluate the resulting encoders using the same downstream full FT protocols across all eight datasets (Values are BAcc (%) and Macro BAcc is the average across the eight datasets).

  1. No visible alignment (w/o boundary loss).
  2. Uniform weighting (w/o boundary-aware weighting).
  3. Boundary-aware weighting (GeoLPE).

Cross-subject LOSO full FT and within-subject full FT results:

Cross-subject LOSO full FT

DatasetGeoLPEw/o boundary-aware weightingw/o boundary loss
BCIC-IV-2a56.1354.8254.71
BCIC-IV-2b76.3871.0871.66
BNCI201500166.0866.0165.06
BNCI201400869.3568.2067.51
BNCI201400978.7276.8476.77
CastillosCVEP10088.1987.7287.89
Hinss202153.2852.5552.64
Kojima2024A69.4468.3166.64
Macro BAcc69.7068.1967.86

Within-subject full FT

DatasetGeoLPEw/o boundary-aware weightingw/o boundary loss
BCIC-IV-2a48.5446.5345.84
BCIC-IV-2b72.8071.8570.35
BNCI201500161.259.3758.39
BNCI201400870.6768.0167.65
BNCI201400977.2575.5973.09
CastillosCVEP10066.6764.5363.37
Hinss202149.4347.7646.58
Kojima2024A63.2462.6562.79
Macro BAcc63.7362.0461.01

With the total alignment weight matched per record, the gains over uniform weighting indicate that the distribution of supervision contributes to the effectiveness of visible-token alignment. These results support boundary-aware weighting as an integral part of the proposed boundary loss, emphasizing visible positions near mask boundaries.

We have added this matched boundary-weighting control, including the control settings and results, to the Experiments section and the appendix of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer 1 (Part 2/4)

Weakness 2: Cross-dataset transfer

The evidence for improved transfer across electrode layouts remains indirect. The geometry ablations in Table 4 establish usefulness within the evaluated settings, but LOSO primarily tests participant transfer, while comparisons across datasets jointly change the task, signal distribution, and montage.

Response to W2

In our additional cross-dataset evaluations, GeoLPE achieves the highest BAcc on all five held-out MI datasets and all three target datasets in the multitask evaluation, with the encoder and readout fixed during target evaluation. The corresponding macro BAcc scores are 64.41% and 56.40%, compared with 61.08% and 52.76% for the strongest baseline, REVE.

Our original manuscript does not claim direct transfer to a different dataset or electrode layout with both the source-trained encoder and readout kept fixed. Our claims specifically concern the generalization of pretrained EEG representations under within-subject and cross-subject LOSO evaluation, with both full fine-tuning and linear probing, and are supported by results across eight datasets. We agree that these original protocols do not isolate robustness to electrode-layout changes, which we examine directly through the same-recording electrode-removal evaluation in our response to Question 1. Directly applying a source-trained encoder and classification head to a held-out dataset without target-specific updates is a more stringent setting beyond the scope of our original evaluation. To address the reviewer’s concern, we additionally evaluated this setting through a five-dataset leave-one-dataset-out MI experiment and a jointly trained multitask model evaluated across three EEG task categories.

1. Task-controlled transfer across five motor-imagery datasets

For downstream cross-dataset evaluation, we harmonized all five datasets to left-hand versus right-hand motor-imagery classification. For BCIC-IV-2a and Stieger2021, we retained only the corresponding left-hand and right-hand MI trials.

In each fold, we used four datasets for supervised source training and held out the remaining dataset for target evaluation, repeating this procedure across all five targets. After source training, we applied the source-trained encoder and classification head directly to the held-out dataset, keeping all parameters fixed and performing no target-dataset fine-tuning or classifier refitting. This evaluates cross-dataset transfer with a common task and label space.

Held-out target datasetGeolpeREVECbramodLabram
bcic-iv-2a67.0164.9151.9362.54
bcic-iv-2b58.9953.7552.9951.54
dreyer202362.8556.2553.4556.13
stieger202169.0766.6362.3350.14
shu-mi64.1263.8859.2356.78
Macro BAcc (%)64.4161.0855.9955.43

GeoLPE achieves the highest BAcc on all five held-out datasets. These results support the transfer of a source-trained encoder and classification head across datasets while holding the task definition and label semantics fixed.

AuthorsOpenReview ↗

Response to Reviewer 1 (Part 3/4)

2. Cross-dataset transfer across three task categories

To examine transfer beyond MI, we jointly trained a single model with a shared encoder and three task-specific classification heads on OpenBMI[1], BI2014a[2], and SEED-SD[3] for motor imagery, P300 detection, and emotion recognition, respectively. We retained the original class definitions in both source and target datasets: left-hand and right-hand MI, target versus non-target P300 responses, and four emotion classes (happy, sad, fear, and neutral). Each training minibatch contained samples from a single source dataset and used the corresponding task-specific head. GeoLPE and all baselines followed this multitask protocol with full fine-tuning of the encoder and classification heads on the source datasets. We then directly evaluated each model on Dreyer2023[4], BNCI2014009[5], and SEED-IV[6], using its source-trained MI, P300, and emotion heads, respectively. All encoder and head parameters remained fixed during target evaluation, with no target-dataset fine-tuning or classifier refitting.

TaskHeld-out target datasetGeoLPEREVECBraModLaBraM
MIDreyer202365.5462.4352.9764.28
EmotionSEED-IV37.7435.5527.8828.78
P300BNCI201400965.9360.3049.1155.76
Macro BAcc (%)56.4052.7643.3249.61

These additional results extend the evidence from downstream adaptation to direct cross-dataset transfer with fixed source-trained encoders and heads, providing further support for the transferability of GeoLPE’s representations. The cross-dataset experiments assess transfer under dataset shifts, while the same-recording electrode-removal evaluation directly examines robustness to changes in electrode coverage with the encoder and readout fixed.

References

[1] Lee M H ,et al. EEG dataset and OpenBMI toolbox for three BCI paradigms: An investigation into BCI illiteracy.

[2] Korczowski L , et al. Brain Invaders calibration-less P300-based BCI using dry EEG electrodes Dataset (bi2014a)

[3] Li Z , et al. Investigating the effects of sleep conditions on emotion responses with eeg signals and eye movements[J]. IEEE Transactions on Affective Computing, 2025, 16(4): 3198-3214.

[4] Dreyer et al., 2023. A large EEG database with users’ profile information for motor imagery brain-computer interface research.

[5] Aricò et al., 2014. Influence of P300 latency jitter on event related potential-based brain-computer interface performance.

[6] Zheng et al., 2019. EmotionMeter: A Multimodal Framework for Recognizing Human Emotions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Common Concerns

Dear Reviewers,

We sincerely thank all reviewers for their constructive feedback and recognition of our work. We would like to clarify our original claims and highlight the additional evidence addressing the shared concerns.

[C1] Geometric Contribution and Cross-Layout Generalization

We thank reviewers KgGt and He1q for highlighting the distinction between supporting heterogeneous electrode montages and reducing performance degradation under layout changes.

Our central geometric contribution is to jointly encode absolute spatiotemporal positions and relative inter-electrode geometry within a unified continuous coordinate system. Our original claim is to support heterogeneous electrode montages through a shared pretrained encoder and dataset-specific downstream adaptation, rather than zero-shot cross-layout generalization of an already fine-tuned model or specifically reducing layout-shift degradation.

Spatial RoPE incorporates relative geometry into cross-channel feature aggregation, so its benefits naturally extend to unchanged layouts (A→A). The observed A→A improvements are therefore consistent with its intended purpose. To provide additional evidence, we conducted the more demanding zero-shot cross-layout evaluation (A→B). Across five seeds, GeoLPE outperforms the variant without spatial RoPE by 3.43 and 2.77 percentage points under A→A and A→B, respectively. These results support the effectiveness of geometry-conditioned aggregation in both settings.

[C2] Methodological Contribution and Integration with Latent Prediction

We thank reviewers ntCL and KgGt for their questions regarding methodological novelty and the contributions of EEG-specific adaptations.

GeoLPE integrates geometric modeling into both feature aggregation and the predictive learning task, rather than simply combining independent positional encoding and JEPA modules. Continuous electrode geometry guides spatial attention, spatial-cluster masking, predictor neighborhoods, masked prediction queries, and boundary-aware alignment.

We further isolated these contributions through geometry-component ablations, positional-mapping comparisons, and matched boundary-weighting controls. The fixed Fourier and learnable coordinate mappings provide complementary multiscale priors and adaptive positional representations. Moreover, boundary-aware weighting improves macro BAcc over matched uniform weighting by 1.51 and 1.69 percentage points under LOSO and within-subject evaluation, respectively, demonstrating benefits beyond additional visible-token supervision.

[C3] Generalization Claims and Supporting Evidence

Our original generalization claim concerns the reusability of pretrained EEG representations across heterogeneous datasets through dataset-specific adaptation, evaluated under LOSO and within-subject protocols with full fine-tuning and linear probing.

To further examine generalization beyond this original scope, we conducted direct cross-dataset transfer and controlled electrode-layout evaluations. In five-dataset zero-shot MI transfer, GeoLPE achieves 64.41% macro BAcc versus 61.08% for REVE. In the same-recording electrode-removal evaluation, GeoLPE also exhibits the smallest performance decrease among the evaluated methods. We additionally clarified dataset provenance and performed trial-level overlap checks.

Matched ECoG controls further isolate the contribution of EEG pretraining, with the pretrained full system outperforming random initialization across all nine participants. To strengthen statistical reporting, we evaluated all ten geometry configurations and the ECoG controls across five random seeds, reporting mean ± standard deviation.

Detailed results and experimental protocols are provided in our individual responses.

As the discussion period approaches its end, with less than one day remaining, we would be grateful if you could consider these additional results and clarifications. We sincerely appreciate your time, thoughtful feedback, and careful consideration of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.