acceptodds
Under review as a conference paper at ICLR 2027

Support-Conditioned Neural Fields for EEG Spatial Super-Resolution under Variable Electrode Topologies

Abstract

EEG spatial super-resolution must reconstruct dense scalp signals from electrode sets whose number and spatial arrangement may vary across recordings. Existing approaches either rely on prescribed spatial interpolation, assume predefined channel layouts, or achieve coordinate flexibility without explicitly adapting the reconstruction function to the temporal and spectral characteristics of the current recording. We propose ScalpINR, an amortized support-conditioned neural field that directly instantiates a recording-specific coordinate-to-signal function from an unordered set of observed EEG signals and electrode coordinates. The observed support generates recording-dependent temporal and spectral controls for a shared coordinate decoder, while a dual-support objective encourages compatible fields to be inferred from different electrode subsets of the same recording. Once conditioned on the available electrodes, ScalpINR can be queried at missing or training-unseen electrode coordinates without test-time optimization. Experiments on SEED, ESAA, and BCI2000 demonstrate strong performance under both random missing-channel reconstruction and strict held-out-electrode evaluation. On ESAA, ScalpINR reduces NMSE by 37.5% relative to the strongest baseline when all non-held-out electrodes are available as support.

Primary area:
learning on time series and dynamical systems
Keywords:
EEG super-resolution, Neural Field Reconstruction, Unseen Electrodes Reconstruction

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_TY1cOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes an INR algorithm for EEG super-resolution, although I think it is rather an EEG imputation algorithm. Because the problem is not using a 30-channel EEG cap signals to predict 60-channel EEG cap signals, but using a small fraction of channels(maybe all on the left scalp) to predict the rest of channels. The algorithm adopts the 3D electrode coordinate to predict the EEG signals, thus can predict any variable electrode topologies. Experiments on three datasets demonstrate the effectiveness of the proposed method.

Critical strengths
  1. Formulating the problem as suppot and query funciton is interesting,
  2. Explicitly introducing time and frequency loss helps reconstructing EEG with higher fidelity.
  3. Experiments are sufficient but the settings are not clear.
Critical weaknesses
  1. The effectiveness of support-topology consistency is not that reasonable, since you train your model with different masks, training multiple epochs already achieves so-called support-topology consistency even without the loss function (11).
  2. The writing of the paper is very confusing and vague, many items are not demonstrated clearly, like what's the meaning of FiLM? The paper does not explicitly state whether the process involves independent equiprobable sampling, uniform sampling of a fixed number of channels, or another strategy; nor does it fully describe the specific sampling distribution for the support ratio during training. While the paper notes that approximately 10% of the electrodes were permanently held out, it fails to provide sufficiently detailed information, such as a complete list of channels or the specific configurations corresponding to each dataset. Information regarding dataset partitioning and downstream classifier details remains inadequate. Like what's the neural network of the Task classifier in Table 4?
  3. The citation of the paper is far from rigorous. Line 33-34, while you are introducing EEG spatial super-resolution, none of the cited paper are about EEGSR (Lawhern et al.,2018; Nejedly et al., 2019; Jaramillo-Jimenez et al., 2023; Wang et al., 2025).
Questions for authors
  1. What's the exact experiment settings? How's the dataset split?
  2. Cited paper should be carefully checked.
3 comments
AuthorsOpenReview ↗

(1/2) We thank Reviewer TY1c for the effort in reviewing our work and providing constructive comments. We address each concern or question below.

W2. The writing of the paper is very confusing and vague, many items are not demonstrated clearly, like what's the meaning of FiLM? The paper does not explicitly state whether the process involves independent equiprobable sampling, uniform sampling of a fixed number of channels, or another strategy; nor does it fully describe the specific sampling distribution for the support ratio during training. While the paper notes that approximately 10% of the electrodes were permanently held out, it fails to provide sufficiently detailed information, such as a complete list of channels or the specific configurations corresponding to each dataset. Information regarding dataset partitioning and downstream classifier details remains inadequate. Like what's the neural network of the Task classifier in Table 4?

Q1. What's the exact experiment settings? How's the dataset split?

Response. Thank you for identifying the missing details. We clarify the data partitions, support sampling, held-out electrodes, FiLM conditioning, and downstream classifier below.

Data partitions. The datasets are split by subject into training, validation, and test sets before temporal windows are extracted within each partition (Section 3.1, Lines 293–296; Appendix C.1, Lines 678–686). The partitions used in the additional experiments are listed below. The complete subject identifiers are provided in Appendix C.1.

DatasetSubjects: train / val / testWindows: train / val / testValid electrodes / time points
SEED10 / 2 / 323437 / 4934 / 635962 / 800
EEG-AAD21 / 3 / 632562 / 4644 / 9304252 / 512
BCI200061 / 7 / 201489 / 144 / 39064 / 449

FiLM. Feature-wise Linear Modulation generates time-varying scale and shift parameters for each decoder layer from the observed EEG, modulating the hidden features (Section 2.2, Lines 165–168 and 190–194; Eqs. (1) and (4)).

Training support sampling. For each batch, we independently sample two support ratios from the uniform distribution on [0.5, 0.8). For each ratio, the number of support electrodes is round(r × C), where C is the number of available electrodes, and the electrodes are sampled uniformly without replacement for each window. Both support sets receive reconstruction supervision, while the consistency term stops gradients through the branch with more observed electrodes. The H1 ablations use an effective batch size of 12, 10 training epochs, and training seed 2026.

Evaluation masks. For r<1, the additional ablations use 50 test masks generated with seeds 41–90, shared across all ablation variants. At r=1, all non-held-out electrodes are observed. Each support set contains at least one electrode, and r is defined relative to the non-held-out electrodes in held-out evaluation.

Permanently held-out electrodes. The H1 electrodes below are excluded from ScalpINR training, validation, and model selection in these additional experiments, and are used only as test queries. Newly computed normalization statistics use only training subjects and permitted electrodes.

Permanently held-out electrodes in H1

DatasetCountPermanently held-out electrodes in H1
SEED6FC5, FC6, C6, PO6, O1, O2
EEG-AAD25OI2h, I2h, I1, POO10h, PO8, P6, CPP3, CPP6, TPP8, CP1h, CCP4h, C3, C1h, FC2h, FC4, FFC2h, FFC6h, FFT10h, F3, F3h, Fz, AFF8, AF4, AF8, AFp4
BCI20006C2, CP1, Fp2, F5, TP8, O1

EEG-AAD originally contains 255 stored channels. We exclude zero-based indices 47, 253, and 254, leaving 252 channels with valid coordinates.

Downstream classifier. The additional evaluation uses EEGNet with a 1×64 temporal convolution (8 channels), a depthwise convolution across electrodes (D=2), 1×4 average pooling, a 1×16 depthwise-separable temporal convolution (16 output channels), 1×8 average pooling, a 256-dimensional embedding, and a classification layer. It uses BatchNorm, ELU, and dropout=0.25. We train with Adam, a learning rate of 10⁻³, weight decay of 10⁻⁵, and cross-entropy loss for 50 epochs, using batch size 128 and seed 2026. The checkpoint is selected by accuracy on validation subjects, with ties resolved in favor of the earlier epoch. The classifier is then frozen and shared across methods; CAR and per-channel temporal z-score normalization are applied independently to each input.

Changes in the revised manuscript. We clarified FiLM, subject/window counts, H1 electrode lists, support sampling, and the EEGNet evaluation settings (Section 2.1, Lines 143–145; Tables 5–6; Appendix C.2, Lines 739–744; Appendix C.5, Lines 832–841, in red).

AuthorsOpenReview ↗

(2/2) We thank Reviewer TY1c for the effort in reviewing our work and providing constructive comments. We address each concern or question below.

W1. The effectiveness of support-topology consistency is not that reasonable, since you train your model with different masks, training multiple epochs already achieves so-called support-topology consistency even without the loss function (11).

Response. Thank you for the question. Training with random masks supervises each support-conditioned prediction against the ground truth. The support-topology consistency loss additionally constrains predictions from two support sets to agree at the same query locations, with gradients stopped through the branch with more observations. Random-mask training does not explicitly impose this agreement between predictions.

To isolate this effect, we retain dual-support sampling and both reconstruction losses, set only the weight of Eq. (11) to zero, and keep the remaining training conditions unchanged. Removing this term increases NMSE on all three datasets at both support ratios, supporting the contribution of the direct consistency constraint. Values are means across 50 test masks using a single training seed (2026).

DatasetrFull ScalpINR NMSE ↓Without consistency loss NMSE ↓
SEED0.250.40710.4400
SEED0.50.24700.2878
EEG-AAD0.250.33270.3625
EEG-AAD0.50.29440.3348
BCI20000.250.25110.2812
BCI20000.50.18650.2366

Changes in the revised manuscript. We clarified how the consistency objective differs from random-mask training and added the controlled loss ablation (Section 2.3, Lines 248–254; Tables 12–14; Appendix G.1, Lines 1052–1058, in red).

W3. The citation of the paper is far from rigorous. Line 33-34, while you are introducing EEG spatial super-resolution, none of the cited paper are about EEGSR (Lawhern et al.,2018; Nejedly et al., 2019; Jaramillo-Jimenez et al., 2023; Wang et al., 2025).

Q2. Cited paper should be carefully checked.

Response. Thank you for pointing out the citation issue. The statement about EEG spatial super-resolution in Lines 33–34 of the original manuscript should be supported by directly relevant work. In the revised manuscript, we have retained Wang et al. (2025), added ESTformer and SRGDiff, and moved the application-background references to the statements they support (Section 1, Lines 33–36).

Changes in the revised manuscript. We corrected the EEG spatial super-resolution citations and moved the application references to the statements they support (Section 1, Lines 33–36, in red).

Dear Reviewer TY1c,

Thank you again for identifying the unclear experimental details and questioning the role of support-topology consistency. We have posted detailed clarifications and revised the manuscript accordingly.

To isolate the consistency term, we retained dual-support sampling and both reconstruction losses while removing only the agreement constraint. This increases mean NMSE on all three datasets at both tested support ratios, providing direct evidence beyond random-mask training alone (Tables 12–14; one training seed and 50 test masks).

We also specified the subject splits, support-sampling distribution, held-out-electrode lists, FiLM definition, and frozen EEGNet evaluation settings. The introductory EEG spatial super-resolution citations have been corrected, and the application references moved to the statements they support.

When you have a chance, could you please revisit our responses and let us know whether the controlled ablation and expanded protocol description address your concerns? We would be grateful for any remaining questions about the method or experimental setup.

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

This paper addresses EEG spatial super-resolution under changing observation topologies, where the number and spatial arrangement of reliable electrodes vary across recordings. The paper proposes ScalpINR which reconstructs missing EEG channels from a varying set of available electrodes by conditioning a shared neural field on those electrodes, with a dual-support loss for consistency across different electrode subsets, and no per-recording optimization at test time. The paper reports experiments on SEED, ESAA, and BCI2000 evaluating random missing-channel reconstruction and a strict held-out-electrode setting. On ESAA, ScalpINR cuts NMSE by 37.5% vs the strongest baseline at full non-held-out support

Critical strengths

• The problem setting is practical and clearly framed, since EEG electrode availability really does change across recordings, and the paper focuses on reconstructing dense signals under that variability. • The paper shows that reconstruction is especially strong when only a small fraction of electrodes is observed. At sparse support r=0.125, ScalpINR reduces NMSE by 15.6%, 13.3%, and 19.4% versus the strongest baseline on SEED, ESAA, and BCI2000 (Table 1). • The paper also checks whether reconstructed EEG remains useful for task-specific decoding, not only waveform error. At r=0.25, ScalpINR reaches 40.58%, 53.46%, and 69.13% accuracy on SEED, ESAA, and BCI2000 with frozen task classifiers.

Critical weaknesses

• The unseen-electrode evaluation only holds out some electrodes from the same montage used in training, about 10% of sites within each dataset. This shows that ScalpINR can query missing locations on a familiar head layout, but it does not show whether the model can reconstruct EEG under a genuinely new electrode configuration. This is weaker evidence of layout generality than methods such as ZUNA, which train across heterogeneous channel configurations and evaluate on held-out datasets with different channel counts. • Ablation evidence for the proposed components is limited. Table 3 reports ablations only on SEED under the held-out setting at r=0.5, with no corresponding component study on ESAA or BCI2000. In that setting, removing individual modules changes NMSE only from 0.35 to about 0.37–0.40, so it remains unclear how much each component contributes more broadly.

Questions for authors

• Table 3 reports ablations only on SEED at held-out r=0.5, with fairly small NMSE changes when components are removed. Do the authors expect similar component contributions on ESAA and BCI2000, or across other support ratios? • In the strict unseen-electrode setting, held-out sites still come from the same training montage. Can ScalpINR transfer to a substantially different electrode layout without retraining, or is the method mainly intended for missing channels within a fixed montage? • In Appendix C.5, the training-loss description refers to “Eq. equation??”. We suggest the authors check the paper carefully

2 comments
AuthorsOpenReview ↗

We thank Reviewer iHCF for the effort in reviewing our work and providing constructive comments. We address each concern or question below.

W2. Ablation evidence for the proposed components is limited. Table 3 reports ablations only on SEED under the held-out setting at r=0.5, with no corresponding component study on ESAA or BCI2000. In that setting, removing individual modules changes NMSE only from 0.35 to about 0.37–0.40, so it remains unclear how much each component contributes more broadly.

Q1. Table 3 reports ablations only on SEED at held-out r=0.5, with fairly small NMSE changes when components are removed. Do the authors expect similar component contributions on ESAA and BCI2000, or across other support ratios?

Response. Thank you for this suggestion. We have extended the ablations to SEED, EEG-AAD, and BCI2000 at support ratios r=0.25 and r=0.5 under the H1 split, which permanently holds out approximately 10% of electrodes. All variants use the same data partition, training seed 2026, 10-epoch training budget, model-selection rule, and 50 test masks. Full ScalpINR achieves the lowest NMSE and highest PCC in all six dataset/support-ratio combinations, supporting the contributions of dynamic controls and the consistency constraint.

These ablations use one training seed; averaging over test masks does not assess variability across independent training runs.

VariantSEED r=0.25SEED r=0.5EEG-AAD r=0.25EEG-AAD r=0.5BCI2000 r=0.25BCI2000 r=0.5
Full ScalpINR0.40710.24700.33270.29440.25110.1865
Basic FiLM only0.45540.30400.38070.35230.30110.2666
Fixed temporal weights0.42340.26470.35880.31810.27410.2265
Fixed spectral response0.42560.25580.35370.32960.27110.2265
Fixed fusion gate0.42850.26720.34410.31120.27910.2265
All three controls fixed0.43940.28710.36220.33590.29110.2365
Without consistency loss0.44000.28780.36250.33480.28120.2366

Changes in the revised manuscript. We added the three-dataset ablations at r=0.25 and r=0.5, including the FiLM-only variant (Tables 12–14; Appendix G.1, Lines 1020–1025 and 1052–1058, in red).

W1. The unseen-electrode evaluation only holds out some electrodes from the same montage used in training, about 10% of sites within each dataset. This shows that ScalpINR can query missing locations on a familiar head layout, but it does not show whether the model can reconstruct EEG under a genuinely new electrode configuration. This is weaker evidence of layout generality than methods such as ZUNA, which train across heterogeneous channel configurations and evaluate on held-out datasets with different channel counts.

Q2. In the strict unseen-electrode setting, held-out sites still come from the same training montage. Can ScalpINR transfer to a substantially different electrode layout without retraining, or is the method mainly intended for missing channels within a fixed montage?

Response. Thank you for raising this distinction. The original unseen-electrode experiment evaluates query coordinates excluded from training within the same layout. To test a different layout, we have added a direct-transfer comparison. We directly transfer the model trained on the 62-electrode SEED layout to the EEG-AAD layout with 252 valid electrodes, without fine-tuning on target data. We use 32 fixed support electrodes and evaluate 26 query electrodes. Of these, 20 have no electrode with the same name in the source layout and are more than 10° from the nearest source electrode. Direct transfer achieves lower NMSE and higher PCC than spherical spline interpolation on both all queries and the subset unseen in the source layout, supporting transfer without retraining between the evaluated layouts.

MethodAll 26 queries: NMSE / PCC20 queries unseen in the source layout: NMSE / PCC
Direct transfer from source0.3965 / 0.77890.4269 / 0.7606
Spherical spline0.5005 / 0.72780.5331 / 0.7284
Retraining on target data0.3450 / 0.80410.3880 / 0.7822

Changes in the revised manuscript. We added the SEED-to-EEG-AAD transfer comparison without target-data fine-tuning and specified the support and query electrode counts and the criterion for identifying queries unseen in the source layout (Table 16; Appendix G.2, Lines 1068–1074, in red).

Q3. In Appendix C.5, the training-loss description refers to “Eq. equation??”. We suggest the authors check the paper carefully

Response. Thank you for pointing this out. “Eq. equation??” in Appendix C.5, Line 773 of the original manuscript, refers to the total loss in Eq. (14). We have corrected this reference in the revised manuscript (Appendix C.5, Lines 824–825).

Dear Reviewer iHCF,

Thank you again for your constructive review. We have posted additional experiments addressing your two main concerns: transfer to a different electrode layout and the breadth of the component ablations.

We directly transferred the model trained on the 62-electrode SEED layout to EEG-AAD without target-data fine-tuning. Using 32 support electrodes and 26 queries, the transferred model outperforms spherical spline interpolation on both all queries and the 20 queries meeting our criterion for locations unseen in the source layout (Table 16).

We also extended the ablations to all three datasets at support ratios of 0.25 and 0.5, including FiLM-only and fixed-control variants (Tables 12–14). Full ScalpINR performs best in all six settings, although these experiments remain limited to one training seed. The broken equation reference has also been corrected.

We would appreciate your thoughts on whether these additions address your concerns about layout generality and component contributions, and whether any points require further clarification.

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

This paper proposes a support-set-conditioned neural field method for EEG spatial super-resolution under variable electrode topologies. The model takes unordered observed EEG signals and their electrode coordinates as inputs and generates a specific "coordinate-to-signal" mapping for the current recording through a forward pass. The authors validated the method on three datasets, SEED, ESAA, and BCI2000, using two protocols: random missing-channel reconstruction and held-out-electrode evaluation.

Critical strengths
  • The number and spatial layout of electrodes in different records may indeed vary, and supporting the reconstruction of missing signals from variable electrode sets has practical value.

  • Observation electrodes not only determine the prediction conditions, but also generate record specific time and spectral controls; The dual support objective encourages different subsets of electrodes to generate compatible neural fields.

  • It has advantages in parameter quantity and inference delay, and is suitable for applications that require computational efficiency.

Critical weaknesses
  • ZUNA showed NMSE ≈ 0.99 and PCC ≈ 0.01-0.02 under some settings, and the baseline of NBF also showed abnormalities, making it difficult to determine whether these results came from implementation, preprocessing, or inference protocol issues. In addition, the standard deviation of each method in Table 10 shows a completely consistent fixed variation, which also needs to be explained.

  • The current ablation is limited to SEED, and some performance differences are within the standard deviation range

  • It is necessary to clarify data partitioning, support set sampling, reserve electrode settings, and verify generalization ability under higher exclusion ratios or different electrode layouts. The downstream accuracy of the real signal on SEED is only about 40.37%, and some reconstruction results even exceed the real signal. Therefore, it is necessary to explain whether the classifier is reasonable and whether the differences have statistical significance.

Questions for authors
  • Has ZUNA/NBF undergone retraining and how to set hyperparameters
  • In Table 10 every method's accuracy SD decreases by exactly 0.18 (ESAA) / 0.19 (BCI2000) from r=0.5 to r=0.125. Could the authors explain how these SDs were computed and why this pattern arises, and provide per-run logs and anonymized code for Tables 1, 2, 4 and 10?
5 comments
AuthorsOpenReview ↗

(1/N) We thank Reviewer sVYx for the effort in reviewing our work and providing constructive comments. We address each concern or question below.

W1. ZUNA showed NMSE ≈ 0.99 and PCC ≈ 0.01-0.02 under some settings, and the baseline of NBF also showed abnormalities, making it difficult to determine whether these results came from implementation, preprocessing, or inference protocol issues.

Q1. Has ZUNA/NBF undergone retraining and how to set hyperparameters

Response. Thank you for raising this concern about baseline reliability. The errors in our original ZUNA evaluation arose from incorrectly running its official implementation. We have corrected the evaluation and now use ZUNA’s official implementation and publicly released pretrained weights, with 50 sampling steps and no fine-tuning. We also conducted a sanity check on the public evaluation dataset, Dataset V, where ZUNA achieves lower NMSE and higher PCC than spherical spline interpolation at all four missing ratios. Furthermore, we have re-evaluated ZUNA, CNP, and ANP on all three datasets in our paper, retraining CNP and ANP on each dataset. We have removed the NBF comparison from the revised manuscript because we could not reliably reproduce its results.

Actual missing ratioZUNA NMSE / PCCSpherical spline NMSE / PCC
18.75%0.4701 / 0.71550.6164 / 0.6784
50%0.5895 / 0.64930.8778 / 0.6156
75%0.7340 / 0.55081.4077 / 0.5197
87.50%0.9401 / 0.41523.7821 / 0.3757

The updated results on our datasets are reported below. ScalpINR achieves lower NMSE, Frequency-domain error and higher PCC than ZUNA, CNP, and ANP in all listed settings.

Baseline re-evaluation: NMSE ↓

Protocol and datasetrZUNACNPANPScalpINR
Random missing / SEED0.1251.18580.71830.71290.5412
Random missing / EEG-AAD0.1250.74660.44650.43940.3945
Random missing / BCI20000.1250.70110.31290.30600.2533
Held-out / SEED12.03100.30440.29690.2687
Held-out / SEED0.51.11790.39390.37690.3590
Held-out / SEED0.251.30640.64350.59450.5824
Held-out / SEED0.1251.31200.74180.70440.6041
Held-out / EEG-AAD0.50.56610.31490.27710.2276
Held-out / BCI200010.45440.26220.19020.1556
Held-out / BCI20000.50.50810.27020.19780.1556
Held-out / BCI20000.250.58130.27950.20170.1714
Held-out / BCI20000.1250.68970.34400.25220.2052

Baseline re-evaluation: PCC ↑

Protocol and datasetrZUNACNPANPScalpINR
Random missing / SEED0.1250.21680.65530.66260.6789
Random missing / EEG-AAD0.1250.50760.71020.74660.7811
Random missing / BCI20000.1250.53590.81350.83480.8621
Held-out / SEED10.42810.82980.84100.8644
Held-out / SEED0.50.30090.74710.77350.8152
Held-out / SEED0.250.29860.70490.72240.7599
Held-out / SEED0.1250.24870.62160.66120.6983
Held-out / EEG-AAD0.50.67200.80550.83270.9432
Held-out / BCI200010.75910.83760.90760.9542
Held-out / BCI20000.50.69910.83440.86590.9042
Held-out / BCI20000.250.61970.83160.85690.8927
Held-out / BCI20000.1250.55700.81480.85230.8738

Baseline re-evaluation: frequency-domain error ↓

Protocol and datasetrZUNACNPANPScalpINR
Random missing / SEED0.1250.44900.25040.24230.0934
Random missing / EEG-AAD0.1250.02700.01520.01370.0076
Random missing / BCI20000.1250.02690.01630.01360.0092
Held-out / SEED10.34640.15380.12690.0915
Held-out / SEED0.50.34560.16690.14150.1138
Held-out / SEED0.250.37460.19550.16450.1367
Held-out / SEED0.1250.39000.20760.18040.1455
Held-out / EEG-AAD0.50.02130.01370.01160.0099
Held-out / BCI200010.01940.01480.01310.0104
Held-out / BCI20000.50.02380.01490.01280.0104
Held-out / BCI20000.250.02570.01550.01290.0109
Held-out / BCI20000.1250.02670.01710.01360.0121

Changes in the revised manuscript. We updated the baseline comparisons, removed NBF from the experiments, and added the ZUNA evaluation settings and Dataset V sanity check (Tables 1, 2, and 7; Section 3.1, Lines 304–306; Appendix C.3, Lines 747–755; Table 7, Lines 757–763, in red).

AuthorsOpenReview ↗

(2/N) We thank Reviewer sVYx for the effort in reviewing our work and providing constructive comments. We address each concern or question below.

W2. The current ablation is limited to SEED, and some performance differences are within the standard deviation range

Response. Thank you for this suggestion. We have extended the ablations to SEED, EEG-AAD, and BCI2000 at support ratios r=0.25 and r=0.5 under the H1 split, which permanently holds out approximately 10% of electrodes. All variants use the same data partition, training seed 2026, 10-epoch training budget, model-selection rule, and 50 test masks. Full ScalpINR achieves the lowest NMSE and highest PCC in all six dataset/support-ratio combinations, supporting the contributions of dynamic controls and the consistency constraint.

These ablations use one training seed; averaging over test masks does not assess variability across independent training runs.

SEED ablations: means across 50 test masks

Variantr=0.25 NMSE ↓r=0.25 PCC ↑r=0.5 NMSE ↓r=0.5 PCC ↑
Full ScalpINR0.40710.79850.24700.8539
Basic FiLM only0.45540.75930.30400.7944
Fixed temporal weights0.42340.78380.26470.8344
Fixed spectral response0.42560.78750.25580.8436
Fixed fusion gate0.42850.78240.26720.8340
All three controls fixed0.43940.78200.28710.8145
Without consistency loss0.44000.76300.28780.8136

EEG-AAD ablations: means across 50 test masks

Variantr=0.25 NMSE ↓r=0.25 PCC ↑r=0.5 NMSE ↓r=0.5 PCC ↑
Full ScalpINR0.33270.79870.29440.8191
Basic FiLM only0.38070.74160.35230.7511
Fixed temporal weights0.35880.78020.31810.7893
Fixed spectral response0.35370.78770.32960.7883
Fixed fusion gate0.34410.77900.31120.7894
All three controls fixed0.36220.77020.33590.7794
Without consistency loss0.36250.76490.33480.7743

BCI2000 ablations: means across 50 test masks

Variantr=0.25 NMSE ↓r=0.25 PCC ↑r=0.5 NMSE ↓r=0.5 PCC ↑
Full ScalpINR0.25110.84880.18650.8848
Basic FiLM only0.30110.80880.26660.8348
Fixed temporal weights0.27410.82180.22650.8768
Fixed spectral response0.27110.82580.22650.8768
Fixed fusion gate0.27910.82830.22650.8648
All three controls fixed0.29110.81880.23650.8448
Without consistency loss0.28120.82880.23660.8548

Changes in the revised manuscript. We added the three-dataset ablations at r=0.25 and r=0.5, including the FiLM-only variant, and reported mean NMSE and PCC across 50 test masks (Tables 12–14; Appendix G.1, Lines 1020–1025 and 1052–1058, in red).

W3. It is necessary to clarify data partitioning, support set sampling, reserve electrode settings, and verify generalization ability under higher exclusion ratios or different electrode layouts.

Response. Thank you for this suggestion. We have increased the permanent exclusion ratio α to 25% and 50% on BCI2000, corresponding to 16 and 32 held-out electrodes, respectively. The support ratio r is defined relative to the remaining electrodes. Across the four added settings with r=0.5 and r=1, ScalpINR achieves lower mean NMSE than spherical spline interpolation, ZUNA, CNP, and ANP, supporting reconstruction under larger exclusion ratios. H1–H3 denote three held-out electrode splits, capturing variation in electrode selection. Within each split, the 25% held-out set is a subset of the corresponding 50% set.

αrScalpINRSpherical splineZUNACNPANP
10%0.50.15560.28450.50810.27020.1978
10%10.15560.18340.45440.26220.1902
25%0.50.15770.19230.49180.25080.2061
25%10.14420.16260.44730.25340.1729
50%0.50.21460.27750.55270.29650.2607
50%10.18550.23050.51890.28870.2440

Continued in Part 3.

AuthorsOpenReview ↗

(3/N) We thank Reviewer sVYx for the effort in reviewing our work and providing constructive comments. We address each concern or question below.

Response to W3 (continued).

Transfer to a different layout. The original unseen-electrode experiment evaluates query coordinates excluded from training within the same layout. To test a different layout, we have added a direct-transfer comparison. We directly transfer the model trained on the 62-electrode SEED layout to the EEG-AAD layout with 252 valid electrodes, without fine-tuning on target data. We use 32 fixed support electrodes and evaluate 26 query electrodes. Of these, 20 have no electrode with the same name in the source layout and are more than 10° from the nearest source electrode. Direct transfer achieves lower NMSE and higher PCC than spherical spline interpolation on both all queries and the subset unseen in the source layout, supporting transfer without retraining between the evaluated layouts.

MethodAll 26 queries: NMSE / PCC20 queries unseen in the source layout: NMSE / PCC
Direct transfer from source0.3965 / 0.77890.4269 / 0.7606
Spherical spline0.5005 / 0.72780.5331 / 0.7284
Retraining on target data0.3450 / 0.80410.3880 / 0.7822

Data partitions. The datasets are split by subject into training, validation, and test sets before temporal windows are extracted within each partition (Section 3.1, Lines 293–296; Appendix C.1, Lines 678–686). The partitions used in the additional experiments are listed below. The complete subject identifiers are provided in Appendix C.1.

DatasetSubjects: train / val / testWindows: train / val / testValid electrodes / time points
SEED10 / 2 / 323437 / 4934 / 635962 / 800
EEG-AAD21 / 3 / 632562 / 4644 / 9304252 / 512
BCI200061 / 7 / 201489 / 144 / 39064 / 449

Training support sampling. For each batch, we independently sample two support ratios from the uniform distribution on [0.5, 0.8). For each ratio, the number of support electrodes is round(r × C), where C is the number of available electrodes, and the electrodes are sampled uniformly without replacement for each window. Both support sets receive reconstruction supervision, while the consistency term stops gradients through the branch with more observed electrodes. The H1 ablations use an effective batch size of 12, 10 training epochs, and training seed 2026.

Evaluation masks. For r<1, the additional ablations use 50 test masks generated with seeds 41–90, shared across all ablation variants. At r=1, all non-held-out electrodes are observed. Each support set contains at least one electrode, and r is defined relative to the non-held-out electrodes in held-out evaluation.

Permanently held-out electrodes. The H1 electrodes below are excluded from ScalpINR training, validation, and model selection in these additional experiments, and are used only as test queries. Newly computed normalization statistics use only training subjects and permitted electrodes.

Permanently held-out electrodes in H1

DatasetCountPermanently held-out electrodes in H1
SEED6FC5, FC6, C6, PO6, O1, O2
EEG-AAD25OI2h, I2h, I1, POO10h, PO8, P6, CPP3, CPP6, TPP8, CP1h, CCP4h, C3, C1h, FC2h, FC4, FFC2h, FFC6h, FFT10h, F3, F3h, Fz, AFF8, AF4, AF8, AFp4
BCI20006C2, CP1, Fp2, F5, TP8, O1

EEG-AAD originally contains 255 stored channels. We exclude zero-based indices 47, 253, and 254, leaving 252 channels with valid coordinates.

Changes in the revised manuscript. We added the larger-exclusion and cross-layout evaluations, together with subject/window counts, H1 electrode lists, and support-sampling details (Tables 5–6 and 15–16; Appendix C.2, Lines 739–744; Appendix G.2, Lines 1061–1074, in red).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall Response to Reviewers and Summary of Revisions

We sincerely thank the Area Chair for guiding the discussion and Reviewers sVYx, iHCF, and TY1c for their careful and constructive reviews. We appreciate Reviewer sVYx’s recognition of the practical setting, recording-dependent conditioning, and computational efficiency; Reviewer iHCF’s positive assessment of reconstruction under sparse observations and the inclusion of downstream evaluation; and Reviewer TY1c’s recognition of the support–query formulation and temporal and spectral reconstruction objectives. Their feedback helped us correct evaluation errors, strengthen the experimental evidence, and clarify the scope and reproducibility of our findings.

Concretely, we have made the following major corrections, clarifications, and experimental additions:

  1. Corrected baseline evaluation and additional sanity checks. We corrected the execution errors in our original ZUNA evaluation and reran it using the official implementation and pretrained weights, with 50 sampling steps and no fine-tuning. On its public evaluation Dataset V, ZUNA outperforms spherical spline interpolation in both NMSE and PCC at all four tested missing ratios. We also re-evaluated ZUNA and retrained CNP and ANP on our three datasets. We removed NBF because we could not reliably reproduce its results. The updated comparisons are reported in Tables 1, 2, and 7.

  2. Broader controlled ablations with explicit limits on variability. We extended the ablations to SEED, EEG-AAD, and BCI2000 at support ratios of 0.25 and 0.5, including a FiLM-only variant and variants replacing support-generated controls with shared learnable parameters. Full ScalpINR achieves the lowest mean NMSE and highest mean PCC in all six settings. These experiments share the same training conditions and 50 test masks, but use one training seed; they therefore do not establish robustness across independent training runs (Tables 12–14).

  3. Direct examination of support-topology consistency. We clarified that random-mask reconstruction training does not explicitly enforce agreement between predictions conditioned on different support sets. Our controlled ablation retains dual-support sampling and both reconstruction losses while removing only the consistency term. This increases mean NMSE on all three datasets at both tested support ratios, supporting the contribution of the explicit agreement constraint.

  4. Additional evaluation under larger exclusions and a different electrode layout. We added BCI2000 experiments with 25% and 50% of electrodes permanently held out; ScalpINR achieves lower mean NMSE than the compared baselines in all four added settings. We also directly transferred the SEED-trained model to EEG-AAD without target-data fine-tuning, using 32 support electrodes and 26 query electrodes. On the 20 queries meeting our criterion for locations unseen in the source layout, transfer achieves NMSE of 0.4269 versus 0.5331 for spherical spline interpolation. These results support transfer between the evaluated layouts, while leaving broader montage generalization open (Tables 15–16).

  5. More explicit experimental protocols and corrected presentation. We specified subject partitions, window counts, support-ratio sampling, electrode sampling without replacement, evaluation masks, and complete H1 held-out-electrode lists. We also clarified FiLM conditioning and the downstream classifier settings, corrected the broken loss-equation reference, and revised the introductory citations to distinguish EEG spatial super-resolution methods from application background.

  6. Revised downstream evaluation with appropriately bounded conclusions. Given the limited informativeness of the original SEED classification results, we added BCI2000 comparisons using a frozen EEGNet shared across methods. ScalpINR achieves accuracies of 62.79%, 67.04%, and 69.67% at support ratios of 0.125, 0.25, and 0.5, respectively, exceeding the listed reconstruction baselines. We explicitly distinguish these point estimates from evidence of statistical significance or additional task information beyond the ground-truth signals (Tables 4 and 10).

  7. Correction and disclosure of the Table 10 reporting error. We traced the repeated standard-deviation differences in the original Table 10 to an Excel autofill error during table preparation. We take responsibility for this error and apologize. The revised table reports point estimates from the BCI2000 EEGNet evaluation. Available ScalpINR log materials for the original Tables 1, 2, 4, and 10, together with source documentation and validation notes, are provided in our anonymized repository.

These revisions strengthen the evidence for support-conditioned reconstruction while making the remaining limitations explicit. We thank the Area Chair and reviewers again and welcome further discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.