acceptodds
Under review as a conference paper at ICLR 2027

NeuroX-EEG: Scalp Field Inpainting for Zero-Shot EEG Super-Resolution

Abstract

EEG electrodes sparsely sample a smooth spatial field on the scalp surface. Existing learning-based super-resolution methods typically reconstruct a fixed set of target channels repeatedly supervised during training, leaving their ability to recover unseen electrode locations unclear. We introduce NeuroX-EEG for zero-shot reconstruction at locations excluded from both conditioning and supervision, a protocol that tests whether reconstruction generalizes over scalp position. Unlike dataset-level zero-shot transfer, this protocol holds out electrode sites within the same recordings. NeuroX-EEG performs diffusion-based inpainting on a 2D scalp grid, where convolution exploits the local smoothness implied by volume conduction, and incorporates pretrained EEG foundation-model representations to supply broader temporal, spectral, and cross-channel structure. Across three datasets spanning 62- and 19-channel montages and multiple observation densities, NeuroX-EEG achieves the lowest NMSE among the evaluated methods under our zero-shot protocol in six of seven dataset–scale settings. Region-wise zero-shot evaluation shows the clearest gains for band-shaped temporal and parietal holdouts, and the reconstructed signals yield higher mean downstream brain–computer interface (BCI) decoding performance than low-density input under the evaluated protocol.

Primary area:
generative models
Keywords:
Electroencephalography (EEG), Spatial Super-Resolution, Scalp-Field Reconstruction, Diffusion Models, Zero-Shot Channel Reconstruction

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

This paper formulates EEG spatial super-resolution (EEGSR) as scalp-field inpainting and utilizes a fixed set of unseen electrodes UU: these specific locations are strictly excluded from training conditions, reconstruction loss calculations, and validation model selection, and are only evaluated during testing. The paper attempts to demonstrate that the model can extrapolate from observed scalp fields to unsupervised electrode sites.

Critical strengths

Fixing the unseen set UU and evaluating exclusively on UU is a significantly stronger and more rigorous approach than the traditional random missing-channel protocol. The experimental problem definition is well-defined and clear. The experiments are relatively comprehensive, covering multiple datasets, different sparsity ratios, and various regional occlusion scenarios.

Critical weaknesses

The most valuable contribution of this paper is its evaluation protocol, not the proposed network methodology. Treating unseen electrodes as field inpainting is conceptually sound, but the paper fails to clearly demonstrate that its specific architecture provides a novel, irreplaceable mechanism compared to existing coordinate-conditioned neural fields, continuous interpolation techniques, or simple field-completion models.

The "zero-shot unseen site" capability is still confined within fixed standard montages or discrete grids. It does not demonstrate true generalization to continuous coordinates, different electrode caps, different coordinate systems, or genuine cross-hardware montages. Therefore, the conclusion should be accurately restricted to "fixed location holdout within the same discrete montage."

Baseline fairness remains highly problematic. Some baseline methods are evaluated in a frozen state, some are mixed-adapted, and others can only be run under different settings. While such comparisons can serve as a reference, they are insufficient to support strong claims regarding state-of-the-art performance.

The quantity, spatial distribution, and distance to the nearest training electrodes for the holdout set UU are limited. The observed performance gains may be heavily influenced by the specific geometric difficulty of the chosen UU rather than representing a genuine site generalization capability.

Questions for authors

Compared to existing coordinate-conditioned neural fields or continuous field completion methods, what crucial new capability does the proposed network mechanism (rather than the testing protocol) actually introduce? It is reasonable to describe EEG reconstruction as scalp-field inpainting, but it leads me to a question: Can the EEG foundation model really produce broader temporal, spectral, and cross-channel structure? What the eeg foundation model gives should be semantic representation, right?

Please retrain adaptable coordinate-query baselines under the exact same fixed UU protocol, and clearly distinguish between a model being "unable to adapt" and simply yielding "poor performance."

Please disclose the electrode identities, random seeds, scalp regions, and distances to the nearest training electrodes for all UU sets, and report performance across multiple independent UU configurations.

Can the model be trained on one montage and tested on a completely different real-world montage, or tested on continuously queried coordinates? If not, are you willing to narrow the scope of the "unseen electrode generalization" claim in the text?

6 comments

We thank the reviewer for the rigorous review and for recognizing the significance in fixing the held-out set and evaluating exclusively with "a significantly stronger and more rigorous approach than the traditional random missing-channel protocol." The review also raises an important point: what the architecture contributes should be shown separately from what the protocol contributes. We agree that making this separation highly strengthens the paper, and we ran additional experiments to this end. The manuscript has been updated accordingly, with all changes marked in blue.

[W1, Q1 (first part), Q2] "... the paper fails to clearly demonstrate that its specific architecture provides a novel, irreplaceable mechanism compared to existing coordinate-conditioned neural fields ..." / "... what crucial new capability does the proposed network mechanism (rather than the testing protocol) actually introduce?" / "Please retrain adaptable coordinate-query baselines under the exact same fixed [U] protocol, and clearly distinguish between ... 'unable to adapt' and ... 'poor performance.'"

To isolate the contribution of the architecture, we trained 88 new models: two recent coordinate-conditioned models, ZUNA1.1 in three training regimes and ScalpINR, each trained on all 22 SEED and SEED-V configurations (4 x 22). The first is ZUNA1.1 (Warner et al., 2026), the updated release of ZUNA, which conditions on continuous 3D electrode coordinates. We trained it in three regimes: fine-tuned from its released weights, and from scratch with its original recipe and with a modified recipe. The second is ScalpINR (Liu et al., 2026), a coordinate-conditioned implicit neural representation for unseen-electrode generation, which we trained from scratch. Every model uses the same held-out electrode set U as NeuroX-EEG. U is fixed for each configuration and never enters the conditioning input or the training loss (Sec. 3.1); for ZUNA1.1, only the seen electrodes are tokenized during training, so the coordinates of U are never presented. Subject splits and metrics are also identical.

NMSE on U, lower is better; bold marks the lowest NMSE among the from-scratch models (PCC and SNR are in the revised Tables 1 and 12):

MethodSEED 2x / 4x / 8xSEED-V 2x / 4x / 8x
Starting from released pretrained weights
ZUNA 1.0, frozen1.039 / 1.061 / 1.0621.000 / 0.999 / 1.000
ZUNA1.1, frozen0.992 / 1.016 / 1.0330.928 / 1.074 / 1.126
ZUNA1.1, fine-tuned0.896 / 0.946 / 0.9840.551 / 0.652 / 0.833
Trained from scratch on the target dataset
ZUNA1.1, original recipe1.018 / 1.005 / 1.0600.939 / 0.982 / 1.058
ZUNA1.1, modified recipe0.932 / 0.998 / 1.0070.672 / 0.758 / 0.997
ScalpINR0.845 / 0.925 / 0.9960.694 / 0.790 / 0.902
NeuroX-EEG0.833 / 0.912 / 1.0330.604 / 0.677 / 0.868

For a fair comparison of architectures, we compare models within the same training regime. When all models are trained from scratch on the target dataset, NeuroX-EEG is the strongest overall: it has lower NMSE than ZUNA1.1 in 5 of 6 settings with the modified recipe and in 6 of 6 with the original recipe. Model size ablation did not narrow this gap. On the configuration where we varied the size of ZUNA1.1 (the recipe ablation, SEED-V 4x C1), a larger model did not improve: a 382M variant that trains well reaches NMSE 0.753 and the 51M modified recipe 0.763, both above NeuroX-EEG (134M) at 0.728 (Tables 13 and 15). We attribute the advantage to the inductive bias of the explicit scalp lattice, which builds local neighborhood structure into the architecture, whereas a coordinate-conditioned model must learn spatial regularity from the same data. ZUNA1.1 shows its strength when it starts from weights pretrained on a large external corpus. Fine-tuned, it showed a better NMSE than NeuroX-EEG in 4 of 6 settings (SEED 8x and all SEED-V scales), by 0.025 to 0.053, while NeuroX-EEG still presented more accurately at SEED 2x and 4x.

ScalpINR is a recent model designed for unseen-electrode generation, and it is the strongest from-scratch competitor in our comparison. Because no official code is available, we re-implemented it from the paper (Appendix C.5). It ranks second to NeuroX-EEG at SEED 2x and 4x and is the most accurate from-scratch method at SEED 8x. NeuroX-EEG has lower NMSE than ScalpINR in 5 of 6 settings and in 16 of the 22 individual configurations (Tables 1 and 15).

Our fine-tuning efforts for ZUNA1.1 also revealed valuable insights. On the original recipe, from-scratch training stays at the level of a zero prediction (NMSE about 1, PCC at most 0.01) in all 11 SEED configurations and 9 of the 11 SEED-V configurations. A one-factor ablation on SEED-V 4x shows that a smaller model (382M to 51M parameters) or a higher loss weight on observed tokens (0.1 to 1.0) is each sufficient to resolve this, while longer training alone (10k to 20k steps) is not (Table 13). The modified recipe combines all three changes, so ZUNA1.1 is adaptable and enters the from-scratch comparison in its trained form. No official fine-tuning code is released, so the training code and hyperparameters are replicated to our best effort (Appendix D.4). Methods that cannot be adapted are discussed under [W3].

Overall, these results answer the question of what "new capability" the network mechanism introduces against "the testing protocol." An explicit scalp lattice lets a model trained on a single dataset reconstruct unseen electrodes more accurately than coordinate-conditioned models trained on the same data. We also make sure our claims are fully grounded. The abstract, Sec. 4.1 and the conclusion make the revised claim within the from-scratch regime, and Sec. 4.1 reports the fine-tuned ZUNA1.1 result. Manuscript updates: Tables 1 and 12 (ZUNA1.1 and ScalpINR), Table 13 (recipe ablation), Table 15 (per configuration), Appendices C.5 and D.4, Abstract, Sec. 4.1, Conclusion. We thank the reviewer for suggesting this comparison, which helped shape the contribution further.

[W2, Q4] "The 'zero-shot unseen site' capability is still confined within fixed standard montages or discrete grids. ... the conclusion should be accurately restricted to 'fixed location holdout within the same discrete montage.'" / "... tested on a completely different real-world montage, or tested on continuously queried coordinates? If not, are you willing to narrow the scope ...?"

We thank the reviewer for kindly suggesting precise definitions and scope. We made a substantial effort to clarify the limitations and potentials of the 2D lattice.

The lattice follows an established convention. Placing electrodes on a 2D topographic grid is a standard representation in CNN-based EEG analysis (Bashivan et al., 2016; Thodoroff et al., 2016; Wang et al., 2020; Su et al., 2022), and we use the TorchEEG layout (Zhang et al., 2024) as the specific assignment (Sec. 3.2).

In our model, the lattice is internal to the model and the actual data format (i.e., what the model reads or outputs) remains the same as other baselines, retaining the ease of comparison. The inputs are the recorded signals of the seen electrodes, placed in their cells without spatial interpolation, and the outputs are the signals at the held-out electrodes, read from their cells. Inputs, outputs and metrics are therefore the same electrode signals for every method in Table 1, and the lattice only defines which electrodes are neighbors inside the network. The lattice is also not tied to one montage: the 62-channel SEED/SEED-V montage and the 19-channel Mumtaz2016 montage are placed on the same reference lattice (Appendix C.4).

What the lattice does limit is where the model can be queried. Queriable targets are within the lattice cells, so the model cannot evaluate at arbitrary continuous coordinates in between. Each model is also trained on one dataset, so we have not shown transfer of a trained model to a different montage or cap. Following the reviewer's suggestion, the abstract, introduction, Sec. 3.1 and the conclusion now describe the setting as "fixed-lattice site holdout within the same montage." Manuscript updates: Abstract, Introduction, Sec. 3.1 and Conclusion (scope); Appendix A (method limitations: no continuous queries, no cross-montage transfer established); Appendix B (both as future work).

[W3] "Baseline fairness remains highly problematic. Some baseline methods are evaluated in a frozen state, some are mixed-adapted ... insufficient to support strong claims regarding state-of-the-art performance."

We thank the reviewer for asking that the training regime of every baseline be clear where the comparison is made. The revised Tables 1, 2 and 9 now label each method by its regime, and we added trained baselines wherever a method can be adapted.

The baseline regimes follow from what each method permits. SSI is spline interpolation with no trainable parameters, so it is training-free. ZUNA 1.0 is evaluated with its released weights because its public release contains inference code and weights but no training loop. SRGDiff cannot output the held-out electrodes without supervising them, because its VAE latent is fitted to a fixed channel set; we therefore train it on the seen electrodes and apply SSI to reach U (hybrid). This is the "unable to adapt" case of the question Q2. ZUNA1.1 is the adaptable case: we wrote its training code ourselves and reported it both fine-tuned and from scratch (see [W1, Q1, Q2]). DeepConv, ScalpINR and NeuroX-EEG are trained from scratch on the seen electrodes.

Consistent with the reviewer's point, the paper now overly claim state-of-the-art performance in generalized settings. The headline claims in the abstract, Sec. 4.1 and the conclusion are made within the from-scratch regime, and every table labels each method by its regime. In that regime, where all models learn from the same data under the same protocol, NeuroX-EEG has the lowest NMSE in 5 of the 6 SEED and SEED-V settings. Checkpoints of NeuroX-EEG and the ZUNA1.1 baselines are selected by NMSE on the held-out electrodes of the validation subjects, and test subjects are never used for selection (Appendix C.3). Manuscript updates: regime labels in Tables 1, 2 and 9; Appendix C.5 (note on baseline adaptation); Sec. 4.1 (headline ranking within regime).

Reviewer_oL8mOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces NeuroX-EEG, a diffusion-based framework for zero-shot EEG channel reconstruction. Unlike conventional EEG super-resolution methods that repeatedly supervise a fixed set of target channels, NeuroX-EEG holds out target electrode sites from both conditioning and supervision throughout training, testing whether reconstruction generalizes across scalp positions rather than memorizing a channel mapping. Partial recordings are projected onto a shared 2D scalp grid, turning channel reconstruction into spatial inpainting, where a 3D spatio-temporal UNet with decoupled spatial and temporal convolutions exploits the local smoothness implied by volume conduction. Pretrained EEG foundation-model representations (DIVER-1) are incorporated through training-time representation alignment, supplying broader temporal, spectral, and cross-channel context. Evaluated on SEED, SEED-V (62-channel), and Mumtaz2016 (19-channel) across 2×/4×/8× densities, NeuroX-EEG attains the lowest NMSE in six of seven dataset–scale settings and the best SNR in five, with the clearest gains for band-shaped temporal and parietal holdouts. Reconstructed signals also improve downstream BCI decoding over low-density input. Its main limitation is that gains are demonstrated on a limited set of datasets, montages, and a fixed scalp lattice.

Critical strengths

++The paper is well written and organized.

++The experimental results demonstrate the technology is effective.

Critical weaknesses

We do not have many additional comments, and the following two points are the ones I am particularly concerned about. Moreover, the workload required for revisions and supplementary experiments will be extremely substantial. For this reason, I only put forward these three suggestions:

--This work raises a question for me: whether EEG super-resolution reconstruction is meaningful. This is mainly because the spatial resolution of signals generated by the human cerebral cortex is very low, roughly around 2 cm. I am not quite clear about the practical significance of performing super-resolution reconstruction on EEG electrodes, and what kinds of experiments or application scenarios can provide feedback to verify the necessity of this research.

--The stated research objective of the paper is zero-shot learning, but in practice, the work still relies on the middle-layer features of a large model to guide training or perform fine-tuning. I am not quite clear where the zero-shot property is embodied, and why the pre-trained large model cannot be directly used to complete the task. Besides, why can the intermediate features of the large model serve as supervision guidance, and can its rationality be theoretically proved?

--Experiment 4.2 claims to realize the reconstruction and restoration of the entire brain region. I am curious that although different brain regions are connected, they actually have huge differences. By what mechanism is the information reconstruction and restoration of such a large area achieved?

Questions for authors

See Weakness.
In conclusion, if the authors can satisfactorily address my concerns, I will recommend ‌strong acceptance‌, as this topic is of great significance.

5 comments

We thank the reviewer for the careful assessment and the accurate summary, and for noting that "this topic is of great significance." We also appreciate that the reviewer kept the requests focused with the revision workload in mind. Each of the three questions led to a clarification in the paper, and the question about using a pretrained model directly is now answered by a new experiment. The manuscript has been updated accordingly, with all changes marked in blue.

[W1] "... whether EEG super-resolution reconstruction is meaningful ... the spatial resolution of signals generated by the human cerebral cortex is very low, roughly around 2 cm. I am not quite clear about the practical significance ... and what kinds of experiments or application scenarios can provide feedback to verify the necessity of this research."

This is a great question, which brings us back to the premise of the task. The low spatial resolution the reviewer describes is in fact what makes the task well posed. Volume conduction through the skull and scalp makes the scalp potential spatially smooth, so the signal at an unrecorded site is largely determined by the surrounding electrodes; the same premise underlies spherical spline interpolation (Perrin et al., 1989), which is routinely used to replace bad channels. Super-resolution in our paper therefore means denser spatial sampling of the volume-conducted scalp field. We do not claim to resolve cortical activity at a finer scale than volume conduction allows, and the introduction now states this scope explicitly.

The practical value lies in making recordings with different electrode sets usable together. Studies and datasets select different subsets of the standard 10-20, 10-10 and 10-5 positions, so their channel sets overlap only partially (Sec. 1). Reconstructing the missing sites maps such recordings onto a common channel set, lets previously collected low-density recordings be used with analyses or models that expect a denser montage, and can reduce reliance on high-density caps.

Downstream decoding provides the experimental check the reviewer asks for. Table 3 feeds reconstructed signals to a fixed decoder. Balanced accuracy rises from 0.4118 with the low-density input to 0.4644 at 2x and from 0.4498 to 0.4842 at 4x, against 0.4972 with the full high-density montage, and NeuroX-EEG gives the highest accuracy among the reconstruction methods at both scales. At 8x, the gain over the low-density input is smaller (0.4485 to 0.4640), and SSI, ZUNA 1.0 and DeepConv decode better; standard deviations are in Table 10. Manuscript updates: Introduction (meaning of super-resolution and practical uses); Sec. 4.3 (decoding).

[W2] "... the work still relies on the middle-layer features of a large model to guide training ... I am not quite clear where the zero-shot property is embodied, and why the pre-trained large model cannot be directly used to complete the task. Besides, why can the intermediate features of the large model serve as supervision guidance, and can its rationality be theoretically proved?"

We thank the reviewer for these three questions. We answer them in order: where the zero-shot property lies, why a pretrained model cannot be used directly (which we now test with a new experiment), and how the alignment is justified.

In our paper, zero-shot refers to electrode sites rather than datasets. The target set U is fixed before training, is never used for conditioning or as a training target (zero loss weight, Eq. 4), and is reconstructed only for held-out subjects at evaluation time. To avoid confusion with dataset-level zero-shot transfer, the revision calls this setting "fixed-lattice site holdout within the same montage." The foundation model does not break this property. DIVER-1 is frozen, used only during the first 20% of training, receives the seen channels only, and is not used at inference. SEED, SEED-V and Mumtaz2016 are not in its pretraining corpus (Appendix C.3), and the model also works without it (NMSE 0.846 without alignment vs. 0.833 with it on SEED 2x; Table 4).

On using a pretrained model directly: DIVER-1, as used here, is a representation encoder. It outputs embeddings for the channels it receives and has no path to produce a signal at an electrode absent from its input. A generative foundation model that can be queried at arbitrary coordinates does exist, namely ZUNA, so we tested the reviewer's question with it directly (setup and full results in our response to Reviewer roHf, [W1, Q1 (first part), Q2]). Used with its released weights, ZUNA1.1 has NMSE 0.928 to 1.126, the level of a zero prediction. Fine-tuned under our protocol, it becomes a strong method and has lower NMSE than NeuroX-EEG in 4 of 6 SEED/SEED-V settings, by 0.025 to 0.053, while NeuroX-EEG is more accurate at SEED 2x and 4x. When both are trained from scratch on the target dataset, NeuroX-EEG is more accurate in 5 of 6 settings. A pretrained model therefore cannot perform the task out of the box but can after fine-tuning, and among models trained on the target dataset alone the lattice model is the most accurate in most settings.

On the alignment: the teacher features do not supervise the reconstructed signal. They regularize intermediate features of the diffusion network through an auxiliary alignment loss, while the reconstruction loss is computed on the seen electrodes only. Representation alignment is an empirically established technique for training diffusion models (Yu et al., 2024), and its justification there, as in our setting, is empirical. We do not claim a theoretical guarantee; the revision describes alignment as an auxiliary, empirically motivated objective and reports its measured effect. Manuscript updates: Abstract, Sec. 1 and Sec. 3.1 (terminology); Sec. 3.6 (alignment); Sec. 4.1 and Table 12 (ZUNA1.1).

[W3] "Experiment 4.2 claims to realize the reconstruction and restoration of the entire brain region. ... By what mechanism is the information reconstruction and restoration of such a large area achieved?"

This question helped us state precisely what Sec. 4.2 reconstructs. The region names are scalp sensor-location labels, not cortical parcellations (Appendix C.4), and the task is to reconstruct the scalp potentials at a contiguous group of electrodes from the remaining ones. Because "region" could suggest restoring the activity of a brain area, we have revised the wording of Sec. 4.2 accordingly.

The mechanism is spatial interpolation of the volume-conducted scalp field. Each scalp electrode records a mixture of many cortical sources, so neighboring electrodes share much of their signal even when the cortical areas beneath them differ functionally. The model therefore does not need to infer the activity of a distinct brain area; it needs observed electrodes around the held-out group from which to interpolate the field.

The results follow this mechanism. When the held-out group is flanked by observed electrodes on several sides, as in the band-shaped Temporal and Parietal holdouts, NeuroX-EEG is the most accurate method (NMSE 0.922 and 0.927 on SEED, 0.625 and 0.848 on SEED-V; Table 2). When a polar cap is removed, as in the Frontal and Occipital holdouts, targets near the pole are supported from one side only; NeuroX-EEG's NMSE there stays close to 1 (0.958 to 1.156), and the best method differs between holdouts (see our response to Reviewer C8sZ, [W2, Q2]). Sec. 4.2 now describes the experiment in these terms. Manuscript updates: Sec. 4.2 (wording and interpretation); Appendix C.4 (region labels).

We thank the reviewer again for the encouragement and the constructive review. We hope that these clarifications and the new experiment address the questions, and we would be grateful if the reviewer would consider strengthening their support for the paper. We would be glad to discuss any remaining questions.

Reviewer_C8sZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduce a method which called NeuroX-EEG for spatial super-resolution that recover the EEG at some unknown points. The method use convolution over grid that captures the spatial smoothness. The method also use pre-trained EEG foundation model for representation alignment. They check the method in zero shot manner and been able to achieved higher reconstruction correlation and lower NMSE compare to previous methods.

Critical strengths

Strengths:

The proposed method leverages a 2D structured lattice to apply local spatial convolutional priors directly as an inpainting task

Projecting 3D data onto a regular 2D grid gives the spatial convolution the structure needed to extract meaningful features

Good prior - The proposed method incorporate REPA (representation alignment) and pre trained EEG model which give prior to the bottleneck of the diffusion model

Critical weaknesses

Weaknesses: This task has already been investigated in other papers that were not used as baselines and may have better results than the proposed method: "EMAG: Differentiable 4D Gaussian Mixture Splatting for EEG Spatial Super-Resolution" Lazarovich et al. This paper also mentions other methods that target the same task.

The proposed method's results decrease significantly in peripheral places due to the convolutional kernal. this need to be adreessed

Mapping the scalp from 3D to 2D introduces a distortion, and this needs to be addressed.

Questions for authors

Please address the weaknesses above:

  1. Using more dataset
  2. Solving the problem in peripheral places
  3. Solving the distortion from the mapping from 2D to 3D
  4. Comparing to SOTA papers
6 comments

We deeply appreciate the careful review and for recognizing that the 2D lattice lets us "apply local spatial convolutional priors directly as an inpainting task" and that alignment with a pretrained EEG model provides a "good prior." The review also points us to EMAG and raises questions on peripheral sites and on the 3D-to-2D mapping. In response, we added two coordinate-conditioned baselines trained under our exact protocol (ZUNA1.1 and ScalpINR), discuss EMAG in the related work, and analyze the peripheral and mapping questions in the main text. The manuscript has been updated accordingly, with all changes marked in blue.

[W1, Q4] "This task has already been investigated in other papers that were not used as baselines and may have better results than the proposed method: 'EMAG ...'" / "Comparing to SOTA papers"

We appreciate the pointer to EMAG, which we now cite and discuss in Sec. 2.1. To compare with recent methods wherever a direct comparison is possible, we added two coordinate-conditioned baselines trained under our exact protocol, and we explain below why EMAG's numbers measure a different task.

EMAG and our setting answer different questions. EMAG fits a 4D Gaussian mixture to each subject, computes its training loss on all high-density electrodes, including the ones it is evaluated on, and splits training and test data by trial within a subject. In our protocol, the evaluated electrodes never enter the conditioning input or the training loss, and testing is on held-out subjects. EMAG thus measures how well a subject-specific model reconstructs electrodes it was trained on, for new trials of the same subject; we measure reconstruction at electrode sites never used in training, for new subjects. EMAG's NMSE is therefore a within-subject error on electrodes that were supervised, whereas ours is a cross-subject error on electrodes that never entered training. Because the two are not the same kind of generalization error, EMAG's values, including those reported on SEED, cannot be entered in Table 1, and its code is to be released upon acceptance, so it could not be retrained under our protocol either.

For the comparison with state-of-the-art methods, we therefore compare with methods that can be run under our protocol. Table 1 already includes SRGDiff, which EMAG identifies as "the current SOTA for deep EEG super-resolution," together with DeepConv, spline interpolation (SSI) and the pretrained ZUNA. We added two coordinate-conditioned models that, unlike EMAG, can be trained under exactly our protocol: ZUNA1.1, the updated release of ZUNA, in three regimes (fine-tuned, and from scratch with two recipes), and ScalpINR, a very recent model for unseen-electrode generation, which we re-implemented and trained from scratch. This adds 88 trained models, four model variants on each of the 22 SEED and SEED-V configurations (setup and full table in our response to Reviewer roHf, [W1, Q1 (first part), Q2]). When all models are trained on the target dataset alone, NeuroX-EEG has lower NMSE than ScalpINR in 5 of 6 settings, and than from-scratch ZUNA1.1 in 5 of 6 (modified recipe) and 6 of 6 (original recipe). ZUNA1.1 fine-tuned from weights pretrained on a large external corpus has lower NMSE than NeuroX-EEG in 4 of 6 settings, by 0.025 to 0.053, and NeuroX-EEG remains more accurate at SEED 2x and 4x. The headline claims are made within a training regime, and every table labels each method by its regime. Manuscript updates: Sec. 2.1 (EMAG); Tables 1 and 12 (ZUNA1.1 and ScalpINR); Appendices C.5 and D.4; Sec. 4.1. We thank the reviewer for this suggestion, which led to a substantially stronger set of baselines.

[W2, Q2] "The proposed method's results decrease significantly in peripheral places due to the convolutional kernal ..."

We thank the reviewer for this question, which led us to examine where and why accuracy drops. The analysis indicates that the drop follows the support geometry of the held-out region rather than its position on the lattice, and the revision states this in Sec. 4.2.

Peripheral position alone does not cause the drop. The Temporal holdout removes the lateral edge of the montage on both sides (F7/F8, FT7/FT8, T7/T8, TP7/TP8, P7/P8 and their neighbors; Table 8), so its targets lie at the lateral edge of the montage, where the lattice cells beyond them hold no electrode and convolutional neighborhoods are incomplete (Figure 4). NeuroX-EEG is nevertheless the most accurate method there (NMSE 0.922 on SEED and 0.625 on SEED-V; Table 2), as it is on the band-shaped Parietal holdout (0.927 and 0.848).

What distinguishes the Frontal and Occipital holdouts is that they remove a whole polar cap. Targets near the pole then have observed electrodes on one side only, and reconstruction becomes extrapolation. This limits all methods, not only convolutional ones: on SEED, no method reaches an NMSE below 0.91 on either polar holdout, and the best method differs between them (DeepConv on Frontal, SRGDiff (+SSI) on Occipital). NeuroX-EEG's NMSE on these holdouts is 0.958 to 1.156.

The lattice boundary may add to this one-sided support effect. Spherical or graph-based operators, which remove the boundary, are a direct way to address it and are part of our future work (Appendix B). In this revision, we state the limitation in the main text, and Table 14 now reports, for every scale-based held-out set, the distance from each target to its nearest seen electrode, so the support geometry of each configuration is explicit (see also our response to Reviewer roHf, [W4, Q3]). Manuscript updates: Sec. 4.2 (support geometry and limitation); Table 14 (Appendix D.5); Appendix B.

[W3, Q3] "Mapping the scalp from 3D to 2D introduces a distortion ..."

We thank the reviewer for raising the mapping question. The reviewer's own summary of our strengths states why we use the mapping: projecting onto a regular 2D grid "gives the spatial convolution the structure needed to extract meaningful features." Below we clarify what the mapping affects and what it does not, and the revision states its distortion in Sec. 3.2.

The mapping follows an established convention. Topographic projections and discrete electrode lattices are standard representations for preserving approximate local electrode relationships in EEG analysis (Delorme & Makeig, 2004; Gramfort et al., 2013; Bashivan et al., 2016; Thodoroff et al., 2016; Wang et al., 2020; Su et al., 2022), and we use the TorchEEG layout (Zhang et al., 2024) as the specific cell assignment.

The mapping does not alter the signals or the evaluation. Each electrode is assigned to one cell, and its recorded signal is placed there without spatial interpolation; the reconstructed signal at each held-out electrode is read from its own cell. Inputs, outputs and metrics are therefore the same electrode signals for every method in Table 1, and the 2D layout only determines which electrodes are treated as neighbors inside the network. Any distortion in the layout can thus affect only our own model's accuracy, not the fairness of the comparison.

The approximate topology is useful in practice. Within the same backbone, replacing the grid with a flattened channel sequence raises NMSE from 0.833 to 6.269 on SEED 2x (Table 4), so the neighborhood structure of the layout is what makes the reconstruction work.

The planar layout does approximate the curved scalp, with the largest deviation near the boundary. Removing this approximation requires spherical, graph-based or continuous convolution over measured electrode coordinates, which is the main direction of our future work (Appendix B). Manuscript updates: Sec. 3.2 (convention and distortion); Appendix B.

[Q1] "Using more dataset"

The evaluation spans three datasets with two montages (62 and 19 channels) and two recording contexts: emotion paradigms in SEED and SEED-V, and clinical resting-state EEG of participants with major depressive disorder and healthy controls in Mumtaz2016. It covers seven dataset-scale settings with 22 scale-based configurations on SEED and SEED-V, eight region holdouts, and downstream decoding. During the rebuttal period, we prioritized the comparison requested under W1 and Q4, training two coordinate-conditioned baselines under the same protocol on every configuration (88 models). The dataset coverage is stated in Appendix A, and training a single model across heterogeneous montages and denser acquisition systems is part of our future work (Appendix B).

We thank the reviewer again for the constructive review. We hope that the added baselines and analyses address the concerns, and we would be glad to discuss any remaining questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank Reviewers roHf, C8sZ, and oL8m for their careful, constructive, and insightful reviews. Reviewer roHf kindly suggested to isolate the contribution of the architecture from that of the evaluation protocol, to retrain coordinate-conditioned baselines under the exact same fixed-[U] setting, and to fully disclose all held-out configurations. Reviewer C8sZ raised important improvements regarding missing baselines such as EMAG, the performance drop at peripheral sites, and the 3D-to-2D mapping distortion. Reviewer oL8m asked fundamental questions about the significance of scalp-field super-resolution, the precise definition of the zero-shot property, and the mechanism underlying large-region reconstruction. Collectively, these comments significantly strengthened the manuscript.

Concretely, we have made the following major revisions:

  • Extensive additional experiments for fair and complete comparisons. We conducted 88 additional training runs (ZUNA1.1 in three training regimes + ScalpINR) on all 22 SEED and SEED-V configurations, enabling an apple-to-apple comparison of architectures.
  • Clear separation of architecture vs. protocol contributions. We now explicitly distinguish performance differences arising from the scalp lattice inductive bias versus those arising from the fixed held-out protocol or from large-scale pretraining.
  • Precise claims. We made precise descriptions over our protocol (Abstract, Introduction, Sec. 3.1, Conclusion) as "fixed-lattice site holdout within the same montage", while explicitly stating the limitations towards continuous coordinates and cross-montage transfer.
  • Improved presentation on baseline regimes. Every method in the revised tables is now clearly labeled by its training regime (frozen / fine-tuned / from-scratch / hybrid / training-free).
  • Full disclosure of held-out sets. We provide complete electrode identities, regional composition, and nearest-neighbor distances for every scale-based U (Table 14 + supplementary material), along with per-configuration results (Table 15).
  • Clarified contribution of the foundation model. DIVER-1 is described strictly as a frozen teacher used only for auxiliary representation alignment during the first 20% of training; it is not used at inference and does not break the zero-shot electrode property.
  • Additional analyses addressing peripheral performance (support geometry rather than lattice position), 3D-to-2D mapping distortion, the practical significance of scalp-field super-resolution, and the interpretation of region holdouts.

All changes are marked in blue in the revised manuscript. Detailed responses to each reviewer follow below. We hope these revisions successfully address the raised concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.