acceptodds
Under review as a conference paper at ICLR 2027

PanoMontage: Learning Montage Fields for Arbitrary-Location EEG Super-Resolution

Abstract

High-density electroencephalography (EEG) captures fine-grained spatial patterns of brain activity, supporting the study of neural dynamics and the localization of abnormal activity, but its hardware and preparation requirements limit widespread use. EEG spatial super-resolution offers a way to recover dense scalp activity from sparse recordings, but most existing methods synthesize signals only at predefined electrode locations, preventing flexible reconstruction at arbitrary scalp locations. We introduce PanoMontage, which formulates EEG super-resolution as continuous montage field learning. Given sparse EEG observations, PanoMontage constructs a personalized montage field that combines physics-guided waveform estimates with learned spatial context, and queries this field to reconstruct EEG signals at arbitrary scalp locations. To enable montage field learning across heterogeneous electrode configurations, we develop a dynamic-montage self-supervised pretraining scheme that jointly optimizes field construction and query reconstruction on large-scale unlabeled EEG recordings. After fine-tuning on four downstream datasets, PanoMontage achieves the lowest NMSE in 12 of 14 dataset-density settings, with reductions of up to 36.5% relative to the strongest competing baseline in each setting. Controlled comparisons further show that the advantage of montage field learning increases as observations become sparser, highlighting its importance for accurate arbitrary-location EEG super-resolution.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG Spatial Super-Resolution, Continuous Spatial Representation, Electrode Montage Generalization, Arbitrary-Resolution Signal Synthesis

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_DxPUOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2(−1)
Soundness
3(+1)
Read review
Paper summary

The paper models EEGSR as a continuous montage field: first, an intermediate spatial field is constructed using spherical splines and the learned corrections, and then decoded at any query coordinate. The method design is complete, and the pre-training, structural ablation, cross-density testing and permanent electrode retention experiments are all relatively thorough.

Critical strengths

The two-stage design of "intermediate field + query decoding" has a clear motivation. The comparison with the parameter load matching of the direct decoder is also quite convincing, and the advantage expands under sparse input. Moreover, the experimental setup of this article also goes beyond the traditional one of a fixed input and output super-resolution experimental setup The 37 permanently left electrode experiments of CCEP-COReg are the most valuable experiments in this paper: they explicitly exclude from downstream input, supervision and validation selection, which is significantly stronger than the common miss-Channel protocol of "random occlusion during training and reconstruction during testing".

Critical weaknesses

The argument of "no position seen" is still insufficient to support the strong statement of "arbitrary position generalization". 37 positions are only unsupervised during the downstream fine-tuning stage. The paper acknowledges that these coordinates may have appeared in the TUEG pre-training, and the 192-point support set in the middle may expose its coordinate geometry. Therefore, this is "no electrodes in downstream supervision", rather than the strict "no coordinates throughout the entire learning process". The strict position reservation experiment only compares Spline and ScalpINR. For work that claims arbitrary coordinate queries, at least ZUNA, EMAG or other coordinate conditioning methods should be supplemented, and it should be ensured that they are retrained according to the same permanently reserved protocol. The main table uses a separate validation set amplitude calibration for SRGDiff and gain/bias calibration for ZUNA. The processing itself is acceptable, but different methods have different degrees of freedom in post-processing. The main conclusion should report both the "original output" and the "results after unified calibration rules" simultaneously. This paper relies on TUEG pre-training and fine-tuning each downstream dataset separately. Therefore, PanoMontage is currently more like a statistical reconstruction model adapted to a dataset rather than a universal operator that can be directly incorporated into any EEG analysis process like spherical spline interpolation. Pre-training itself is not a drawback, but the paper has not yet demonstrated that a frozen model can stably outperform classical interpolation on brand-new devices, brand-new montage, and brand-new data distributions. The current performance improvement may partly rely on pre-trained data prior and downstream adaptation.

Questions for authors

Will the 37 reserved coordinates appear in TUEG pre-training, fixed latent support, or any coordinate encoding? Please provide a more rigorous version of "The coordinates of the entire training process have not been seen". Can a completely frozen PanoMontage, which has only been pre-trained by TUEG, stably outperform spherical spline on new datasets, new electrode caps, and new reference methods that have never been used for training or fine-tuning? Please report the results separately for "de Novo Training" and "Dataset Fine-tuning". If fine-tuning is still needed for each new dataset, do the authors agree to position the method as data-adapted EEG reconstruction instead of a general preprocessing operator that can directly replace classical interpolation? Can the model output uncertainty or reject reconstruction prompts for inputs that do not cover the pre-trained distribution? This is particularly important for the application of super-resolution in clinical or downstream neuroscience analysis.

9 comments
AuthorsOpenReview ↗

To Reviewer DxPU (Part I)

We thank you for recognizing the clear motivation of the two-stage intermediate-field design, the capacity-matched direct-decoder comparison, and the value of the permanent 37-electrode holdout experiment. Below we address coordinate exposure, uniform calibration fairness, frozen cross-dataset transfer, and uncertainty.

Q1. Coordinate exposure and comparisons under permanent location exclusion.

Response:

None of the 37 held-out coordinates appeared in TUEG pretraining—not in the observed inputs, intermediate support, supervised queries, or coordinate encodings. Across the TUEG pretraining corpus and coordinate-sampling protocol, pretraining draws intermediate coordinates from the standard 10–20 coordinate pool (Appendix B.4), where none of the 37 locations is sampled as an observed input, intermediate support point, or supervised query, and coordinate encodings are computed directly from the supplied input, support, and query coordinates without any external position-embedding lookup table.

To separate unseen target measurements from unseen target coordinates across the entire pipeline, we evaluate PanoMontage alongside newly added coordinate-queryable baselines (Functional ScalpINR and EMAG) under two distinct protocols. Within each protocol, all methods use identical test windows and observation masks.

(1) Protocol A: Held-out measurements, known downstream coordinates (Original Paper Setting). During downstream fine-tuning and validation on CCEP-COReg, the 37 electrodes provide no input signals and no supervision, though their coordinates remain in the query set (and 34 enter the 192-point intermediate support). Because intermediate waveforms are inferred purely from observed electrodes, coordinate presence supplies spatial geometry rather than target EEG measurements:

DensityPanoMontageSplineFunctional ScalpINREMAG
2×0.20610.21890.21691.3017
4×0.23630.29600.24980.7147
8×0.29740.54150.31210.5587
16×0.39531.03560.41810.5630

(2) Protocol B: Strict coordinate exclusion throughout training (Newly Added). All learned models are trained from random initialization on CCEP-COReg (without TUEG weights), and all 37 target locations plus their 1∘1^\circ spatial neighborhoods are completely removed from training and validation inputs, intermediate support, and queries:

DensityPanoMontageDirectSplineFunctional ScalpINREMAG
2×0.21120.20990.21650.21180.7678
4×0.23940.24500.30070.24660.4855
8×0.29860.32050.54610.30870.4566
16×0.38910.43611.02290.40690.5313

PanoMontage outperforms Functional ScalpINR and EMAG at all four densities under both protocols, and outperforms Direct and Spline at 4×–16× under strict coordinate exclusion.

AuthorsOpenReview ↗

To Reviewer DxPU (Part II)

Q2. Different post-processing freedoms across methods; report raw outputs and uniformly calibrated results.

Response:

PanoMontage’s advantage holds both for raw outputs and when every method is granted the exact same linear post-processing freedom. We conducted a paired evaluation across 8 methods and all 14 dataset–density settings on a shared recording-stratified test panel (distinct from the full-test Table 1 benchmark). For each method, dataset, and density, we evaluate (i) raw standardized-waveform predictions and (ii) predictions after a global affine calibration (gain+bias\text{gain} + \text{bias}) fitted exclusively on validation data with model weights frozen.

PanoMontage achieves the lowest mean NMSE in 13 of 14 settings under both raw and uniformly calibrated evaluations, and outperforms Functional ScalpINR and EMAG in all 14 settings. Each entry below reports raw → calibrated NMSE:

DatasetDensityPanoMontageESTformerFunctional ScalpINREMAG
CCEP-COReg2×0.2096 → 0.20970.2145 → 0.21470.2215 → 0.22190.2857 → 0.2855
CCEP-COReg4×0.2447 → 0.24480.2509 → 0.25080.2539 → 0.25460.3378 → 0.3374
CCEP-COReg8×0.3017 → 0.30200.3132 → 0.31340.3160 → 0.31580.4082 → 0.4080
CCEP-COReg16×0.3980 → 0.39740.3998 → 0.39940.4190 → 0.41570.5098 → 0.5099
Localize-MI2×0.1973 → 0.19690.3271 → 0.31610.2435 → 0.24300.2747 → 0.2753
Localize-MI4×0.2256 → 0.22560.3239 → 0.31950.2658 → 0.26570.3272 → 0.3272
Localize-MI8×0.2822 → 0.28160.3605 → 0.35930.3143 → 0.31410.3895 → 0.3892
Localize-MI16×0.3728 → 0.36730.4321 → 0.43190.3996 → 0.39570.4634 → 0.4633
SEED2×0.2816 → 0.28230.3157 → 0.31490.3062 → 0.30530.3648 → 0.3628
SEED4×0.3378 → 0.33780.3483 → 0.34750.3615 → 0.35860.4551 → 0.4540
SEED8×0.4384 → 0.43490.4340 → 0.43260.4434 → 0.43950.5434 → 0.5421
SEED-IV2×0.1693 → 0.16940.1724 → 0.17260.1796 → 0.17960.2527 → 0.2519
SEED-IV4×0.2290 → 0.22920.2298 → 0.22980.2367 → 0.23640.3341 → 0.3336
SEED-IV8×0.3276 → 0.32550.3334 → 0.33330.3287 → 0.32710.4405 → 0.4394
DatasetDensitySplineSTADTopoDiffSRGDiff
CCEP-COReg2×0.2263 → 0.22350.4399 → 0.36400.7313 → 0.70243.9039 → 0.9297
CCEP-COReg4×0.3070 → 0.29300.4898 → 0.40360.8337 → 0.80214.5782 → 0.8870
CCEP-COReg8×0.5065 → 0.42810.5882 → 0.49080.9318 → 0.92922.7206 → 0.8663
CCEP-COReg16×0.9871 → 0.63340.8399 → 0.65980.9718 → 0.97174.2336 → 0.9668
Localize-MI2×0.3074 → 0.29720.4365 → 0.38750.6293 → 0.62681.7413 → 0.7361
Localize-MI4×0.3531 → 0.33480.4681 → 0.42480.7405 → 0.74251.7462 → 0.7621
Localize-MI8×0.4905 → 0.43560.5199 → 0.45740.9014 → 0.88141.8912 → 0.8463
Localize-MI16×0.8890 → 0.62880.6171 → 0.53520.9736 → 0.95841.6459 → 0.8961
SEED2×0.4531 → 0.41550.5879 → 0.50600.8153 → 0.81152.2178 → 0.5496
SEED4×0.7714 → 0.58300.6517 → 0.55510.9779 → 0.96611.4857 → 0.6603
SEED8×1.4999 → 0.77860.7786 → 0.65520.9855 → 0.97432.4006 → 0.8547
SEED-IV2×0.2270 → 0.21920.3952 → 0.36350.9858 → 0.97316.5202 → 0.5550
SEED-IV4×0.4335 → 0.37620.4500 → 0.40750.9709 → 0.96987.0296 → 0.6426
SEED-IV8×0.9746 → 0.59950.5886 → 0.51091.0152 → 0.99885.5190 → 0.7444
AuthorsOpenReview ↗

To Reviewer DxPU (Part III)

Q3. Dependence on downstream adaptation; frozen transfer to new datasets, caps, and references, and comparison with training from scratch.

Response:

(1) Frozen zero-shot transfer across five datasets. A single frozen TUEG-pretrained PanoMontage checkpoint outperforms spherical spline in 9 of 10 dataset–density settings without any downstream weight updates, checkpoint selection, or fitted output calibration. All inputs use an invertible normalization computed solely from observed channels: for each window, we subtract the instantaneous observed-channel mean and apply a single shared median/MAD-based center and scale to the residuals, without clipping. Predictions are mapped back to the original amplitude scale using these same observed-only quantities; no target-channel statistics are used. To test transfer on a brand-new dataset, we added the PhysioNet EEG Motor Movement/Imagery Dataset (EEGMMIDB) (64 channels; 20 test subjects, all 14 runs/subject covering resting state, executed movement, and motor imagery):

Dataset4× NMSE: PanoMontage / Spline8× NMSE: PanoMontage / Spline
CCEP-COReg0.3549 / 0.32130.4315 / 0.4909
Localize-MI0.5108 / 0.61840.5906 / 0.8454
SEED0.7013 / 1.84650.7459 / 2.0092
SEED-IV0.4657 / 0.49870.5571 / 0.9559
EEGMMIDB (newly added)0.6506 / 0.89330.7102 / 1.4966

On EEGMMIDB, the frozen model reduces NMSE by 27.2% at 4× and 52.5% at 8× over Spline.

(2) De novo training vs. dataset fine-tuning. Appendix E.1 (Table 17) reports the three-seed comparison on CCEP-COReg:

DensityDe novo training (from scratch)Pretraining + fine-tuning
2×0.2230 ± 0.00170.2171 ± 0.0011
4×0.2490 ± 0.00070.2430 ± 0.0001
8×0.3140 ± 0.00130.3079 ± 0.0010
16×0.4161 ± 0.00090.4042 ± 0.0032

(3) Positioning of the method. Together, these results show that PanoMontage is a pretrained continuous-field reconstruction model capable of direct frozen cross-dataset transfer, while supporting optional target-domain adaptation for maximum accuracy.

Reviewer_X6USOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PanoMontage, an EEG super-resolution model designed to reconstruct brain signals at any requested scalp coordinate from a sparse set of recorded electrodes. Rather than predicting missing channels in a single step, the model first builds an intermediate montage field on a fixed grid of points, combining classical spherical-spline interpolation with learned waveform and spatial-context corrections, and then queries that intermediate field to generate the final waveforms. The model is pretrained on unlabeled TUEG data using varying electrode layouts and fine-tuned on four downstream datasets. While PanoMontage achieves the lowest mean NMSE in 12 of 14 test settings and shows clear architectural benefits when input channels are very sparse, the evaluation is limited to independently normalized waveforms rather than true physical scalp voltages.

Critical strengths
  1. Blending classical spline interpolation with learned residual corrections is a sensible design choice, and the authors provide useful evidence that the intermediate stage offers more than additional model capacity. Compared with a direct reconstruction model without an intermediate montage, with nearly identical parameter count and similar computation (Tables 4 and 20), PanoMontage reduces NMSE by approximately 8% at 8× and 16% at 16× on CCEP-COReg. Additional controls on SEED-IV (Table 5) support contributions from both the intermediate waveform estimates and the transmitted context features, although these component comparisons use only one training seed.
  2. Rather than only testing randomly dropped channels on a fixed cap, the authors test unseen sparsity levels (3x, 5x, 10x) and run a strict held-out location experiment on CCEP-COReg where 37 electrodes are completely hidden during downstream training and validation.
Critical weaknesses
  1. The introduction motivates EEG super-resolution through improved spatial characterization of scalp activity and potential applications to source localization. However, preprocessing independently standardizes every channel, including held-out test targets, using its own trial- or run-level median and median absolute deviation (Appendix A.1, Eq. 10). Although target statistics are not supplied to the model, this evaluation does not require recovery of the original channel-specific amplitude scales or the physical voltage relationships between electrodes. Those relationships matter for interpreting scalp topographies and performing source localization. Appendix A.1 acknowledges that restoring physical units uses the target channel’s scale and offset, which would be unavailable at an unrecorded location. The results therefore establish standardized waveform reconstruction, while the connection to the paper’s broader practical motivation remains insufficiently demonstrated.
  2. Although PanoMontage achieves the lowest mean NMSE in 12 of 14 settings, five of the 14 paired confidence intervals include zero: all three SEED-IV densities and CCEP-COReg at 4x and 16x. One interval favors ESTformer, at SEED 8×. The test populations are also small: SEED-IV has two test subjects, who favor different methods at every density, while the largest reported gain comes from Localize-MI, with one test subject and nine recordings. The authors acknowledge these limitations and provide useful uncertainty estimates, but the main comparisons use selected training runs rather than averages over multiple seeds. Consequently, the reliability and broader generalizability of the small advantages over ESTformer remain uncertain.
Questions for authors
  1. Does PanoMontage retain its advantage when using a common normalization derived only from observed channels, preserving relative voltage amplitudes across the scalp? How would predictions at an unrecorded location be converted into physical microvolts without that location’s normalization statistics?
  2. Several key baselines (ScalpINR, ESTformer, STAD, TopoDiff) were re-implemented from paper descriptions rather than official code, and ZUNA was evaluated zero-shot with only a linear validation scale/offset adjustment rather than full target-domain fine-tuning. More concerningly, the reproduced ScalpINR model has 0.983M parameters in Table 22, whereas the original ScalpINR paper reports 2.563M parameters, so my question here is what accounts for the parameter discrepancy in your ScalpINR reproduction (0.983M in Table 22 vs. 2.563M in the original paper)?
  3. How stable are the main performance differences across training seeds and alternative subject splits, particularly for SEED-IV and Localize-MI? Repeating the closest comparisons would help distinguish consistent improvements from training and subject variability.
  4. EMAG is discussed as a closely related method but is absent from the experiments. Why was it excluded, and how does PanoMontage’s scalp-based intermediate representation differ conceptually from EMAG’s source-based space–time Gaussian representation?
2 comments
AuthorsOpenReview ↗

To Reviewer X6US (Part I)

We appreciate your positive assessment of our spline-guided intermediate field, parameter-matched controls, and evaluation at unseen densities and permanently withheld locations. Below we address physical amplitude recovery, baseline implementations, seed/split stability, and EMAG.

Q1. Recovering inter-electrode amplitudes without target-channel statistics, and performance under observed-only common normalization.

Response:

PanoMontage preserves its advantage when normalizing solely with observed-channel statistics and inverting predictions back to the original amplitude scale. While our original per-channel normalization isolates waveform shape, recovering true inter-electrode voltage relationships at unrecorded locations requires a shared, observation-only transformation.

(1) Invertible observed-only normalization. For each window with observed channels O\mathcal O, we compute the instantaneous observed-channel mean reference r(t)r(t), along with a single scalar center mm and scale ss across all observed residuals:

r(t)=mean⁡i∈Oxi(t),zi(t)=xi(t)−r(t)−ms,r(t)=\operatorname{mean}_{i\in\mathcal O}x_i(t),\qquad z_i(t)=\frac{x_i(t)-r(t)-m}{s},

Here, mm and ss are a shared median-based center and MAD-based scale computed solely from observed-channel residuals. At any unrecorded query coordinate qq, the physical-scale waveform is recovered directly as:

x^q(t)=sz^q(t)+m+r(t).\widehat x_q(t)=s\widehat z_q(t)+m+r(t).

Because r(t)r(t), mm, and ss depend only on O\mathcal O and are shared across all channels, relative voltage amplitudes across the scalp are preserved and no target-channel statistics are used.

(2) Evaluation in original amplitude units. We adapted PanoMontage, ESTformer, Functional ScalpINR, EMAG, and SRGDiff (three adaptation seeds each from their fixed checkpoints) under this shared protocol at 4×, inverted their outputs, and evaluated full-test NMSE in native amplitude units (μV\mu\text{V} for CCEP-COReg and Localize-MI; native source units for SEED and SEED-IV):

MethodCCEP-CORegLocalize-MISEEDSEED-IV
PanoMontage0.2584 ± 0.00080.3394 ± 0.00040.5667 ± 0.00050.2612 ± 0.0004
Functional ScalpINR0.2672 ± 0.00020.3982 ± 0.00400.6725 ± 0.00130.2662 ± 0.0018
ESTformer0.2735 ± 0.00030.4959 ± 0.00100.6189 ± 0.00040.2784 ± 0.0018
Spline0.32130.61841.84640.4987
EMAG0.3551 ± 0.00100.4578 ± 0.00330.7306 ± 0.00190.3784 ± 0.0025
SRGDiff0.8842 ± 0.00650.7585 ± 0.00520.6740 ± 0.00780.6395 ± 0.0048

PanoMontage achieves the lowest amplitude-scale NMSE on all four datasets.

AuthorsOpenReview ↗

To Reviewer X6US (Part II)

Q2. Reliability of reimplemented baselines, the ScalpINR parameter discrepancy, and the zero-shot ZUNA setting.

Response:

(1) ScalpINR parameter discrepancy and Functional ScalpINR. The original ScalpINR paper reports 2.563M parameters but did not release official code or layer-by-layer hyperparameter tables. Our initial baseline was an independent implementation from its text description (983,426 parameters), and we cannot attribute the exact parameter gap from the published total alone. We now explicitly label that model as an independent paper-based implementation.

To ensure a rigorous comparison against public author-supplied code, we additionally evaluated Functional ScalpINR using the official supplementary implementation and full configuration from Support-Conditioned Neural Fields for EEG Spatial Super-Resolution under Variable Electrode Topologies (437,884 parameters):

Model / Implementation SourceTrainable Parameters
ScalpINR (reported in original paper; no code released)2.563M
ScalpINR (our independent paper-based implementation)983,426
Functional ScalpINR (official supplementary code & full config)437,884

Trained under identical subject splits, observation densities, and validation checkpoint selection, Functional ScalpINR is substantially stronger than our paper-based ScalpINR implementation on SEED and SEED-IV, yet PanoMontage still achieves lower NMSE in all 14 settings.

(2) ZUNA evaluation setting. The original ZUNA paper releases pretrained weights and inference code specifically for direct zero-shot application to new datasets and does not define a target-domain fine-tuning protocol. We therefore evaluated ZUNA in its published frozen-inference regime with validation-fitted linear calibration.

Q3. Stability across training seeds and alternative subject splits, particularly on SEED-IV and Localize-MI.

Response:

We evaluated seed and subject-split stability:

(1) SEED-IV across three training seeds. We trained PanoMontage and ESTformer across three seeds each, selecting checkpoints on a fixed validation panel and evaluating uncalibrated standardized-waveform NMSE on identical test windows and masks:

DensityPanoMontage NMSE ↓ESTformer NMSE ↓
2×0.1721 ± 0.00100.1768 ± 0.0036
4×0.2247 ± 0.00130.2338 ± 0.0071
8×0.3183 ± 0.00180.3151 ± 0.0020

PanoMontage achieves lower mean NMSE at 2× and 4×, whereas ESTformer is slightly lower at 8× (consistent with the overlapping confidence interval in Appendix D.2, Table 13). Notably, PanoMontage exhibits smaller across-seed standard deviations at all three densities.

(2) Localize-MI across seven held-out subject rotations. Repeating training and evaluation across seven disjoint train/validation/test subject rotations (covering seven distinct held-out test subjects on matched test windows) confirms that our Localize-MI gains are not an artifact of a single test participant:

DensityPanoMontage NMSE ↓ESTformer NMSE ↓Spline NMSE ↓
2×0.1285 ± 0.03620.2216 ± 0.06200.2160 ± 0.0521
4×0.1457 ± 0.04010.2181 ± 0.06020.2473 ± 0.0635
8×0.1819 ± 0.04650.2491 ± 0.06420.3439 ± 0.0915
16×0.2540 ± 0.05510.3225 ± 0.07310.6453 ± 0.1718

Q4. Missing EMAG comparison and the distinction between scalp-based and source-based representations.

Response:

(1) Conceptual and architectural distinction. While both methods reconstruct scalp EEG at continuous coordinates, EMAG models a latent 3D source space via space–time Gaussians and projects their activity onto the scalp through a forward rendering operator. Furthermore, EMAG is formulated as a subject-specific optimization model fitted per recording/subject (EMAG, Section 4.3). By contrast, PanoMontage operates directly on the scalp manifold: it constructs explicit intermediate scalp waveforms X^L\widehat X_L anchored by spherical splines alongside learned spatial context HLH_L, avoiding an ill-posed inverse-source intermediate bottleneck and generalizing directly to held-out subjects in a feed-forward pass.

(2) Empirical comparison with EMAG. Because official EMAG code was unavailable, we implemented EMAG from its published specification and adapted it to our cross-subject benchmark. PanoMontage achieves lower NMSE than our EMAG reproduction across all 14 dataset–density settings, both before and after validation-fitted linear calibration.

Reviewer_ed2mOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an EEG spatial super-resolution as continuous montage field learning to to reconstruct EEG signals at arbitrary scalp locations.

Critical strengths

The manuscript addresses EEG super-resolution by proposing a reconstruction method based on a continuous montage field and evaluating it on multiple datasets. The reported results show low reconstruction errors and suggest potential utility.

Critical weaknesses

The practical value of the task requires clearer justification, and additional experimental evidence is needed to establish the distinctive advantages of the proposed method and support the main conclusions.

Questions for authors
  1. The practical value of EEG super-resolution and the evidence needed to establish it should be clarified. Volume conduction spatially smooths scalp potentials; therefore, increasing the number of reconstructed electrode signals does not directly establish improved spatial resolution of the underlying neural activity. Moreover, scalp EEG contains background brain activity, artifacts, and measurement noise. Low reconstruction errors or high signal correlations alone are insufficient to demonstrate that the model recovers spatial information with neurophysiological relevance. The authors should clarify what useful spatial information the method recovers and how this can be distinguished from fitting noise or reproducing learned data priors. I recommend additional validation tied to the intended application. For example, under the same low-density input conditions and a consistent evaluation protocol, does super-resolution improve downstream recognition or classification performance? Alternatively, can the reconstructed signals improve the analysis of topographic patterns, spatial gradients, or localized responses in signals with well-characterized spatial structure? Such evidence would help establish the engineering or neurophysiological value of the task.

  2. The experimental design does not sufficiently demonstrate the distinctive advantages of the continuous montage field Based on the methodological description, the continuous montage field can be understood as a learnable, adaptive spatial interpolation approach that uses electrode geometry and data priors. This is a reasonable principle, but the current experiments primarily demonstrate overall reconstruction performance without sufficiently establishing the specific advantages arising from this design. The experiments should be more directly motivated by the method’s continuous spatial representation. Relevant settings could include electrode locations unseen during training, irregular electrode layouts, varying observation densities, and transfer across montages. Ablation studies should also distinguish the contributions of geometric information, the continuous field representation, and other network components. Furthermore, the authors should clarify the key methodological differences from existing spatial interpolation and continuous implicit representation approaches, and provide corresponding experimental evidence.

  3. The baseline coverage is limited, and the claimed advantage under increasingly sparse observations is not sufficiently supported The comparisons are limited to ScalpINR and spherical spline interpolation, which provides an insufficient basis for assessing the method’s advantages over existing EEG spatial reconstruction approaches. I recommend including additional representative learning-based reconstruction methods and other applicable spatial interpolation approaches, with consistent observed electrodes, training data, and evaluation protocols. In addition, the trend in Figure 2 appears to indicate that the proposed method deteriorates more rapidly as the observation scale factor increases. This does not directly support the claim that its advantage becomes greater as observations become sparser. The authors should first clarify the relationship between the observation scale factor and the number of observed electrodes, then report absolute performance, performance differences relative to the baselines, and variability across observation densities. The corresponding claim should be reassessed in light of these results.

  4. The representativeness of the datasets and the recovery of spatial details require further investigation. The localized-mi and ccepcoreg datasets involve intracranial electrical stimulation with simultaneous scalp EEG recording. Such stimulation-evoked responses may exhibit relatively localized and pronounced spatiotemporal patterns, and their signal characteristics and reconstruction difficulty may not adequately represent those of conventional spontaneous or task-related EEG. The authors should discuss how this dataset selection affects the scope of their conclusions and, preferably, provide additional validation on data that more closely reflect the intended application. Furthermore, the evaluation mainly relies on NMSE and signal correlation, which may not adequately capture the recovery of local spatial structure. In the topographic maps in Figure 3, SRGDiff appears to recover more local details consistent with the reference maps. The authors should explain this observation and its relationship to the quantitative results. If recovering spatial detail is a central objective of the super-resolution task, the evaluation should also examine features such as spatial peak locations, local topographic patterns, or spatial gradients, rather than relying solely on aggregate reconstruction error and signal correlation.

2 comments
AuthorsOpenReview ↗

To Reviewer ed2m (Part I)

We thank you for your suggestions on spatial-topography validation and methodological positioning. Below we address spatial value and evidence for the intermediate montage field.

Q1. Practical spatial value beyond waveform error: downstream or spatial-structure validation.

Response:

PanoMontage recovers more accurate scalp potential topographies from sparse inputs—mitigating spatial aliasing on the scalp rather than claiming to bypass volume conduction in neural source space. Because sparse electrode arrays undersample scalp potential gradients and local extrema, recovering faithful scalp topographies is essential for topographic mapping.

Following your suggestion to evaluate topographic patterns beyond channel-wise NMSE/PCC, we evaluated Global Map Dissimilarity (GMD) (Brunet et al., 2011, Eq. 2), which measures polarity-sensitive structural disagreement between mean-centered, GFP-normalized scalp maps at unobserved electrodes (lower is better). We evaluate GMD at 4× across all four datasets using offline restoration from original checkpoints, with per-channel normalization statistics on matched test samples:

DatasetPanoMontageESTformerFunctional ScalpINREMAGSplineSRGDiff
CCEP-COReg0.56380.57090.57500.65210.59871.1018
Localize-MI0.63390.76880.69430.76770.77371.1523
SEED0.88890.88990.89040.95420.98321.0749
SEED-IV0.55830.56240.56480.66960.69540.9560

These results support improved scalp-topography recovery across the evaluated datasets. Scalp-map fidelity does not, by itself, establish separation of neural sources from volume-conducted background noise.

Q2. Evidence for the intermediate field beyond adaptive interpolation, including unseen locations and varying observation layouts and densities.

Response:

PanoMontage uses a two-stage decomposition with explicit intermediate waveforms and spatial context.

(1) Methodological distinction. Classical spherical spline applies fixed geometry-only weights (X‾Q=SQ←OXO\overline X_Q = S_{Q\leftarrow O}X_O). Conditional INR methods such as ScalpINR compress sparse observations into a latent conditioning vector and decode each target coordinate independently. In contrast, PanoMontage first constructs an explicit intermediate scalp field FL=(PL,X^L,HL)\mathcal F_L = (P_L, \widehat X_L, H_L) and then queries it (Eq. 2):

X^Q=SQ←LSL←OXO+SQ←LΔL+ΔQ.\widehat X_Q = S_{Q\leftarrow L}S_{L\leftarrow O}X_O + S_{Q\leftarrow L}\Delta_L + \Delta_Q.

Here, explicit intermediate waveforms X^L\widehat X_L propagate first-stage learned corrections ΔL\Delta_L through the second-stage spline operator and attention-based waveform mixing, while spatial context HLH_L transmits high-dimensional observation features. Because both OO and QQ interact with FL\mathcal F_L via continuous coordinates and set-based cross-attention, a single model naturally handles unseen target coordinates, irregular layouts, and varying observation densities.

(2) Evidence for the intermediate field. The manuscript provides controlled comparisons that distinguish the field representation from model capacity and interpolation alone:

  • Intermediate field vs. direct decoding: Section 4.4, Figure 4(a) compares the two-stage model with a parameter-matched single-stage Direct decoder, holding coordinate access and spline guidance constant; Appendix E.3, Table 20 documents the capacity comparison.
  • Intermediate waveforms vs. spatial context: Section 4.4, Figure 4(b) and Appendix E.1, Table 16 separately ablate the learned intermediate waveforms and transmitted context, supporting the contribution of both paths.
  • Analytical geometry vs. learned corrections: Section 4.5, Figure 5 and Appendix E.1, Table 17 compare variants without the spline baseline and with a spline-only constructor, isolating their respective contributions.
  • Changing densities and layouts: Figure 2 evaluates unseen observation densities, while Table 6 and Appendix E.2, Tables 18–19 evaluate dynamic montage sampling. Additional strict coordinate-exclusion and frozen cross-dataset experiments test generalization beyond these settings.

Together, these controls support the intermediate field as a useful representation for reconstruction, beyond simply increasing model capacity or applying fixed spatial interpolation.

AuthorsOpenReview ↗

To Reviewer ed2m (Part II)

Q3. Limited baseline coverage and insufficient evidence for increasing advantage under sparse observations.

Response:

(1) Baseline coverage. Table 1 evaluates Spline and six learning-based baselines: ESTformer, STAD, SRGDiff, TopoDiff, ScalpINR, and ZUNA. We additionally evaluated Functional ScalpINR and EMAG; PanoMontage achieves lower NMSE than both in all 14 dataset–density settings, before and after validation-fitted linear calibration. Figure 2 and Table 2 focus on continuous-density and unseen-coordinate queries, motivating comparisons with coordinate-queryable methods.

(2) Scale factors and the trend in Figure 2. Larger factors denote fewer observed electrodes and generally make reconstruction harder, although individual methods may show non-monotonic error trends. For CCEP-COReg, including Figure 2, NO=⌊Nvalid/r⌋N_O=\lfloor N_{\mathrm{valid}}/r\rfloor for each recording. For SEED/SEED-IV and Localize-MI, factors are nominal density labels: observed-channel counts are 32/16/8 at 2×/4×/8× and 128/64/32/16 at 2×/4×/8×/16×, respectively. Within each dataset and density, all methods use the same observed-channel counts.

A steeper error curve does not itself imply a growing advantage. The relevant comparison is the error difference at each density. The Figure 2 results (Appendix D.3, Table 14) are:

FactorPanoMontage NMSESpline NMSEScalpINR NMSEGain vs. Spline (absolute / relative)Gain vs. ScalpINR (absolute / relative)
2×0.20780.22220.34520.0144 / 6.5%0.1374 / 39.8%
3×0.22270.25450.35360.0318 / 12.5%0.1309 / 37.0%
4×0.23880.29030.36340.0515 / 17.7%0.1246 / 34.3%
5×0.25540.33530.37870.0799 / 23.8%0.1233 / 32.6%
8×0.29520.49750.41330.2023 / 40.7%0.1181 / 28.6%
10×0.31710.60130.43590.2842 / 47.3%0.1188 / 27.3%
16×0.38500.97760.50070.5926 / 60.6%0.1157 / 23.1%

The gain over Spline grows with sparsity, whereas the gain over ScalpINR does not. ScalpINR starts with a higher error at dense observations, which helps explain its flatter curve. The three-seed, parameter-matched control in Section 4.4, Figure 4(a) and Appendix E.3, Table 20 provides separate evidence for the intermediate-field mechanism:

DensityPanoMontage (scratch)Direct decoderRelative gain
2×0.2230 ± 0.00170.2106 ± 0.0004-5.9%
4×0.2490 ± 0.00070.2508 ± 0.00040.7%
8×0.3140 ± 0.00130.3415 ± 0.00118.1%
16×0.4161 ± 0.00090.4950 ± 0.001815.9%

Thus, the increasing sparse-observation advantage is supported relative to Spline and the parameter-matched Direct decoder in these controls, not universally across external model families. For example, the SEED-IV three-seed comparison favors PanoMontage at 2×/4× and slightly favors ESTformer at 8×.

Q4. Representativeness of stimulation datasets and SRGDiff’s apparent advantage in local spatial detail.

Response:

(1) Dataset representativeness. Localize-MI and CCEP-COReg contribute scalp EEG recordings only; no intracranial channels are used in training or evaluation. Stimulation-evoked recordings alone cannot establish performance on conventional EEG. Our benchmark also includes cognitive/emotion EEG from SEED and SEED-IV. We further evaluated frozen transfer on the PhysioNet EEG Motor Movement/Imagery Dataset (EEGMMIDB), covering resting state, motor execution, and motor imagery across 20 test subjects. Frozen PanoMontage reduces NMSE over Spline by 27.2% at 4× and 52.5% at 8×, extending the evidence beyond stimulation-evoked recordings.

(2) Spatial detail in Figure 3(a). Sharper-looking scalp-map variation is not necessarily more accurate. In the reported example, SRGDiff overestimates the right-posterior positive focus and reverses the ground-truth polarity at held-out electrode E190; PanoMontage preserves the negative polarity and local gradient, with snapshot spatial MAE 0.3388 versus 0.9754. This is a local example, while the GMD comparison in Q1 evaluates spatial-topography agreement across four datasets: PanoMontage has lower GMD than SRGDiff on each dataset. Together, these comparisons support more faithful scalp-map reconstruction, rather than interpreting visual sharpness alone as recovered physiological detail.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.