acceptodds
Under review as a conference paper at ICLR 2027

EMAG: Differentiable 4D Gaussian Mixture Splatting for EEG Spatial Super-Resolution

Abstract

High-density electroencephalography (HD-EEG) enables fine-grained measurement of cortical activity but requires expensive hardware and lengthy setup times, limiting its clinical and research accessibility. We propose EMAG (EEG Mixture of Anisotropic Gaussians), a differentiable framework that reconstructs HD-EEG signals from a sparse subset of low-density (LD) electrodes by representing brain electrical sources as a mixture of anisotropic 4D space-time Gaussians. EMAG places a mixture of multiple Gaussians at each point of a spherical brain grid, each parameterized by a full precision matrix, enabling anisotropic spatial spreads and explicit coupling between spatial and temporal dimensions. The forward model renders scalp EEG via differentiable Gaussian field contributions at electrode locations, enabling end-to-end training without explicit source localization supervision. We evaluate EMAG on three public EEG benchmarks (Localize-MI, SEED, and SEED-IV) at super-resolution factors of through . EMAG outperforms the current state-of-the-art EEG super-resolution method at most super-resolution factors. The learned Gaussians carry explicit 3D brain-space coordinates and can be directly visualized; this spatial grounding potentially opens avenues for clinical and neuroscientific applications such as source localization.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG super-resolution, 4D Gaussian mixture, differentiable rendering, spatiotemporal EEG reconstruction, personalized brain modeling, low-density EEG

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

This paper proposes EMAG, representing the spatial super-resolution of EEG as the anisotropic 4D Gaussian mixture on the brain grid. The model is driven by low-density EEG to the temporal amplitudes of each Gaussian component and rendered from the potential source field to the scalp electrode through differentiable Gaussian kernels. The author contends that this representation possesses both reconstruction performance and potential source space interpretability.

Critical strengths

Interpreting EEGSR as a differentiable mapping from a latent source field to electrode signals provides a more structured inductive bias compared to pure electrode-space networks.

The 4×4 SPD precision matrix, the spherical brain mesh, and the low-density (LD) signal modulation collectively form a complete and computationally feasible architecture.

Critical weaknesses

The methodological novelty is significantly overstated. The original 4D Gaussian Splatting already introduced spatiotemporal Gaussian representations, and 4D-Rotor Gaussian Splatting has also utilized native 4D covariance. The primary modification in EMAG is adapting the existing 4D Gaussian representation to EEG using a Cholesky formulation, a brain mesh, and LD-conditioned amplitudes. While these are reasonable domain-specific engineering efforts, they are insufficient to constitute strong methodological innovation.

The so-called "differentiable splatting/forward model" is essentially evaluating and computing a weighted sum of Gaussian/RBF kernels at electrode coordinates. It does not involve the camera projection, ray integration, occlusion, or alpha compositing typically found in computer graphics. More importantly, this kernel does not incorporate the true EEG leadfield, dipole orientations, tissue conductivity, head geometry, or reference schemes. Therefore, it should be accurately positioned as an empirical geometric prior, rather than an authentic EEG forward model or a reliable source imaging technique.

The current experiments fail to rule out the interpretation that the Gaussian representation is merely a more complex basis function for interpolation. Specifically, the full HD montage is supervised during training, and the primary metrics are calculated including observed channels. This neither validates generalization to unseen locations nor prevents the overall error metrics from being heavily skewed by the easily replicated input channels.

The comparisons with strong baselines lack auditability. Baselines need to be rerun under the exact same trial splits, input masks, normalization schemes, training budgets, and random seeds, rather than mixing in previously reported results from other papers.

Questions for authors

Compared to the original 4D-GS and 4D-Rotor GS, which core mechanism is mathematically novel, rather than merely migrating existing 4D covariance or Gaussian rendering techniques to EEG?

If this method indeed learns a meaningful source field, can it be rigorously compared against quantitative localization baselines (e.g., true leadfield/BEM forward models, MNE/sLORETA, or known stimulation sites), instead of solely relying on qualitative visualizations showing Gaussians near stimulated regions?

Please report evaluation metrics computed only on unobserved electrodes, and conduct tests with permanent electrode position holdouts or cross-montage evaluations.

Please rerun all strong baselines under a unified protocol and report the variance and statistical significance.

3 comments

Response (1/2)

We thank the reviewer for the detailed and rigorous review, and for recognizing that a source-field-to-electrode mapping "provides a more structured inductive bias compared to pure electrode-space networks" and forms "a complete and computationally feasible architecture". We address each point below.


W1 / Q1 — What is novel relative to 4D-GS and 4D-Rotor?

A: We agree that the 4D Gaussian primitive is not ours, and we do not claim to have invented Gaussian splatting or 4D Gaussians. We credit 4D-GS and 4D-Rotor and describe their 4D covariance explicitly (Sec. 1–2). Our contribution is the first use of the Gaussian-splatting analogy for EEG, together with the adaptations required to make that analogy work in the EEG domain. Transferring an established technique to a new domain, with the adaptations that domain requires, is a well-recognized form of contribution: the Vision Transformer (Dosovitskiy et al., ICLR 2021) brought the NLP Transformer to images with minimal architectural change, and diffusion-based EEG-SR methods, including SRGDiff, adapt diffusion models from vision. Our adaptations are:

  • Sources anchored in brain space. Gaussian mixtures are placed on a volumetric brain grid rather than optimized freely in the scene. Without this anchor, the centers migrate to the scalp and become shortcut interpolators (App. F.3). Restricting sources to a cortical shell degrades SEED by +116% (Table 2).
  • Electrodes as the "camera". Rendering targets the 3D electrode coordinates of the montage instead of image pixels (Sec. 3.2).
  • Measurement-driven sources. In vision, 4D-GS models dynamics by deforming the scene per frame. In EEG, the sparse LD measurements must drive the sources, so we modulate per-grid-point source amplitudes from the LD input (Sec. 3.4). Replacing this with a single global modulation raises Localize-MI NMSE five-fold (0.105 → 0.541, Table 2).
  • Efficient evaluation at fixed sensors. Because the electrodes and the grid are fixed, the 4D Mahalanobis distance decomposes into precomputed spatial terms plus a cheap per-timestep term (Eq. 4–6), making the full 4×4 covariance practical for long recordings.

Each adaptation is validated by ablation (Table 2, App. F).


W2 — Is the Gaussian rendering a forward model?

A: We agree that EMAG is not a biophysical forward model, and the paper positions it this way. Sec. 1 describes the renderer as "a physics-inspired surrogate that encodes the core volume-conduction inductive bias, a smooth distance-decaying source-to-sensor projection, rather than an exact biophysical head model". We use the term physics-inspired consistently (Sec. 2, 4.4, F.5, 6).

The simplification is deliberate, and its structure matches EEG physics:

  • A weighted sum is the correct form. The EEG forward problem is linear: scalp potentials are a superposition of source contributions weighted by the leadfield. Occlusion and alpha compositing are absent because volume conduction has no occlusion, not because the analogy is loose. EMAG keeps this linear-superposition structure and replaces the leadfield with a learned, smoothly decaying kernel.
  • Gaussian decay as an analogy to volume-conduction decay. We know the Gaussian is not the exact physical kernel. It captures the property that matters for super-resolution, smooth decay of influence with distance, and its learned anisotropy absorbs orientation and conductivity effects that we do not model explicitly.
  • No MRI, conductivity model or head geometry is needed, so the method applies wherever these are unavailable, which is the typical setting for sparse EEG.

We also tested the physically derived alternative directly: replacing the Gaussian kernel with a fixed three-shell spherical leadfield on the same grid performs markedly worse on both datasets (Table 2, App. F.5). Where the paper uses "forward model", it refers to this learned source-to-sensor mapping in the generative sense, not to a biophysical leadfield.

(continued in the next comment)

Response (2/2)

W3 / Q3 — Is EMAG just an interpolation basis? Metrics on unobserved electrodes.

A: Several controlled experiments test the interpolation hypothesis directly:

  • Adding explicit interpolation does not help. Feeding the model a spherical-spline-interpolated HD input instead of the raw LD input gives no improvement (Table 2, App. F.4).
  • A pure interpolator is much worse. Non-learned spherical-spline interpolation of the LD montage performs far worse than EMAG (App. D.1, Table 4).
  • A black-box mapping is worse. A capacity-matched electrode-space MLP, which can represent any interpolation scheme, doubles NMSE on Localize-MI (Table 2).
  • When EMAG is allowed to interpolate, it gets worse. With free-floating centers, components migrate to the scalp next to the LD electrodes and act as local interpolators, and performance drops (App. F.3). The brain-anchored grid is what prevents this.

Reporting over all HD channels follows the convention of SRGDiff and its baselines (App. B), keeping our numbers directly comparable; these baselines do not publish unobserved-electrode metrics, so no like-for-like comparison exists in the literature. As requested, we computed NMSE on the unobserved electrodes only (mean over five split seeds):

Dataset (2× / 4× / 8× / 16×)All channelsUnobserved only
Localize-MI0.087 / 0.105 / 0.140 / 0.2140.105 / 0.123 / 0.157 / 0.227
SEED0.156 / 0.280 / 0.4240.305 / 0.369 / 0.478
SEED-IV0.119 / 0.208 / 0.3400.229 / 0.274 / 0.382

On the dense 256-channel montage the cost of excluding observed electrodes is small, and EMAG's unobserved-only NMSE on Localize-MI remains below SRGDiff's published all-channel NMSE at every factor (0.145 / 0.238 / 0.296 / 0.346). On the 62-channel caps the difference between the two metrics is larger, because a larger fraction of the electrodes is observed (half of them at 2×). This applies to any method under the all-channel convention, including the baselines.


W4 / Q4 — Baseline protocol, variance and statistical significance.

A: We agree that a fully re-run comparison is the ideal, and we took several steps toward it. SRGDiff's code is public, but its exact train/test split and trained weights are not, so its published numbers cannot be reproduced sample for sample. We therefore followed SRGDiff's published protocol as closely as possible: the same datasets, SR factors, random 80/20 split, uniform random LD subsampling and metrics (App. B). The only element we cannot match is their specific random draw of the split. The unavailability of a split file should not prevent comparison with published results; comparing under a matched protocol is standard practice.

To rule out a favorable split, we follow the common practice of using multiple seeds: five independent split seeds per setting (SRGDiff reports a single split). Seed-to-seed variability is very small (std ≤ 0.0024 NMSE in every setting). A one-sample t-test of the five EMAG runs against SRGDiff's published values shows that EMAG's improvement is statistically significant in every setting where it leads, on all three metrics: p < 10⁻⁷ on Localize-MI at every factor, p < 3·10⁻³ on SEED at 2× and 4×, and p < 7·10⁻³ on SEED-IV at 2× and 4×. At 8× on the 62-channel caps, SRGDiff remains ahead, as reported transparently in the paper (Sec. 4.2).

In addition, we evaluated a further strong baseline, the 380M-parameter EEG foundation model ZUNA, entirely under our own protocol (identical splits, LD electrodes, ground truth and metrics). EMAG outperforms it on every dataset and SR factor (e.g. Localize-MI 4×: NMSE 0.105 vs 0.467; SEED 2×: 0.156 vs 2.019).


Q2 — Quantitative comparison of the learned sources.

A: We agree with the reviewer that a rigorous source-localization claim would require such comparisons, and this is why we do not claim that EMAG recovers the true neural sources. EMAG is trained without any source supervision. The paper says only that high-amplitude components "cluster near the known stimulation sites" (Sec. 6). In App. H.1 we explicitly state that this "is not precise localization" and that intracranial stimulation drives network-level propagation, so the scalp signal reflects multiple regions. Our goal there was to check whether the learned representation shows any correspondence with real sources. A rigorous comparison with MNE or sLORETA would require subject-specific head models, beyond the scope of a super-resolution paper.


We are grateful for the reviewer's rigorous feedback, which led to the unobserved-electrode evaluation, the significance tests and the ZUNA baseline. We hope these address the reviewer's main concerns, and would sincerely appreciate it if the reviewer would reconsider the score.

Thank you for the clarifications and additional experiments. I appreciate the unobserved-electrode evaluation and the clarification that EMAG does not recover validated neural sources. However, several concerns remain unresolved. First, evaluating electrodes absent from the test input is different from evaluating positions permanently excluded from training supervision. The added results address the former, but the requested permanent electrode holdout or cross-montage evaluation remains missing. Second, the explanation that the larger metric gap on the 62-channel datasets arises from a larger observed-electrode fraction appears inconsistent with the stated super-resolution factors: at 2×, approximately half the electrodes are observed in both the 62-channel and 256-channel settings. Please clarify this discrepancy. Finally, I still think that what you are doing now is merely applying a mature method to EEG, and you haven't demonstrated where the difficulties lie in its application to EEG, or what difficulty of EEG super-resolution you solve.

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

The paper proposes EMAG for reconstructing high-density EEG from sparse electrodes. It represents a latent field using anisotropic 4D Gaussians on a fixed brain grid, modulates their amplitudes using the input EEG, and sums their contributions at target electrode locations. Experiments evaluate reconstruction on Localize-MI, SEED, and SEED-IV.

Critical strengths
  • The method is simple and provides an explicit, spatially grounded representation for EEG reconstruction.
  • Reconstruction improves on Localize-MI and at lower upsampling factors on SEED and SEED-IV. Ablations support the Gaussian parameterization and input conditioning.
Critical weaknesses
  • Limited evaluation scale. Evaluation covers only three relatively small datasets. Larger-scale EEG benchmarks, such as TUEV (or others), would better establish scalability.
  • Unclear downstream utility. Although Appendix C reports a downstream experiment, the classifier is unspecified. A three-way comparison between low-density inputs, reconstructed signals, and original high-density signals is needed to show the remaining performance gap. It is also unclear whether reconstruction remains beneficial with stronger models, such as EEG foundation models.

Minor:

  • It could be argued that the method is relatively simple, mainly adapting existing 4D Gaussian representations to EEG reconstruction.
  • Given the many implementation and experimental details involved, open-source code is needed to reproduce the reported results.
Questions for authors

See weaknesses.

2 comments

We thank the reviewer for the positive assessment and for recognizing that the method is simple and "provides an explicit, spatially grounded representation", and that the ablations support the design. Following the review, we now provide the full specification of the downstream classifier and the requested three-way comparison (LD / reconstruction / original HD) (W2). We address each point below.


W1 — Evaluation on larger datasets.

A: We recognize that many EEG datasets could be used for super-resolution. We deliberately chose the benchmarks used by the current state of the art for this task (SRGDiff and the baselines it reports), so that our results can be compared directly against published numbers on the same datasets, super-resolution factors and metrics. These benchmarks cover two very different regimes: focal evoked responses on a 256-channel montage (Localize-MI) and continuous affective recordings on a 62-channel cap (SEED, SEED-IV).


W2 — Downstream classifier details and LD / reconstruction / HD comparison.

A: We thank the reviewer for this concrete suggestion. Following common practice in EEG super-resolution and the protocol of SRGDiff, we use a simple, fixed classifier: a random forest (100 trees) on 5-band differential-entropy features, trained per subject on SEED (3 emotion classes) with an identical train/test split for every input type. The three-way comparison is below (accuracy; chance = 0.33):

Input2×4×8×
LD input0.4090.3270.333
EMAG reconstruction0.6100.5340.502
Original HD (upper bound)0.7200.7200.720

EMAG's reconstruction recovers a large part of the gap between the sparse input and the full HD recording (+0.20 / +0.21 / +0.17), consistently for all 15 subjects (paired t-test, p < 10⁻⁸). The original HD signal serves as the upper bound for any reconstruction method. We will add this table and the classifier details to App. C.

Regarding stronger models: solving a downstream task well is a substantial project in its own right, separate from reconstruction. Prior EEG-SR work therefore demonstrates utility with a fixed, simple classifier applied under the same protocol to all inputs, and we follow this protocol.


W3 — The method is relatively simple.

A: We see this simplicity as a strength rather than a weakness. A deliberate goal of EMAG is to simplify EEG super-resolution: an explicit source representation with only ~73k parameters per subject, about 30× fewer than SRGDiff, trained on a single GPU. Yet it outperforms considerably larger diffusion models on most settings. The ablations (Table 2) show that each component of this simple design is necessary.


W4 — Code availability.

A: We fully agree. All code, configurations and experiment scripts will be publicly released upon acceptance. In the meantime, App. A–B give the complete implementation and training details (architecture, initialization, optimizer, schedules, splits, seeds and preprocessing).


We hope that the added downstream details and three-way comparison resolve the reviewer's concerns, and that they strengthen the reviewer's support for the paper. We would be grateful if the reviewer would consider raising the score, and we are happy to provide any further details.

Reviewer_EFxkOpenReview ↗

Thanks

Thank you for the response. My concerns are only partially addressed, particularly regarding larger-scale datasets and downstream evaluation. The current evidence remains limited to establish practical benefits. I therefore maintain my original score.

Reviewer_5W5iOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper propose EMAG (EEG Mixture of Anisotropic Gaussians), a differentiable framework that reconstructs HD-EEG signals from a sparse subset of low-density (LD) electrodes by representing brain electrical sources as a mixture of anisotropic 4D space-time Gaussians

Critical strengths

1.The mathematical formulation is clear, concise, and elegant. The Gaussian representation, amplitude modulation, and differentiable rendering process are presented within a coherent framework, making the method straightforward to understand.

  1. The ablation studies are systematic and informative. They examine key design choices, including Gaussian dimensionality, covariance coupling, spatial constraints, and input conditioning, helping clarify the contributions of individual components.
Critical weaknesses

1.The novelty is limited and the distinctive advantages of the proposed 4D Gaussian representation are not sufficiently established. Anisotropic 4D Gaussian representations with explicit spatiotemporal covariance modeling have already been explored in prior work [1,2], The paper’s novelty appears to lie primarily in adapting and combining these components for EEG reconstruction. The paper would benefit from a clearer explanation of which EEG-specific design improvement is made. In addition, the paper lacks direct performance comparisons with closely related approaches, including Neural Brain Fields (NBF), masked autoencoders trained with whole-channel masking, and ZUNA [3]. Such comparisons in experiment are needed to establish the specific benefits of the proposed representation.

  1. The experimental evidence for the method’s generalizability and practical utility remains limited. The evaluation covers a narrow range of EEG tasks; additional experiments on sleep, motor imagery, and disease-related EEG datasets would strengthen the evidence for broader applicability. Furthermore, the practical benefit of reconstructing high-density EEG requires more convincing downstream validation. The authors should compare their reconstruction-based pipeline with strong models trained directly on the same low-density EEG input. This would clarify whether reconstructing additional channels provides meaningful performance gains over directly training an effective model on the available electrodes.

[1] Duan et al. (2024). 4D-Rotor Gaussian Splatting: Towards Efficient Novel View Synthesis for Dynamic Scenes. ACM SIGGRAPH 2024.

[2] Yang et al. (2024). Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting. ICLR 2024.

[3] ZUNA: Flexible EEG Superresolution with Position-Aware Diffusion Autoencoders

Questions for authors

please see the weakness

2 comments

We thank the reviewer for recognizing that the formulation is "clear, concise, and elegant" and that the ablations are "systematic and informative". Following the reviewer's suggestion, we evaluated ZUNA on all three benchmarks under our exact protocol; EMAG outperforms it in every setting (W2).


W1 — What is novel beyond existing 4D Gaussian splatting?

A: We agree the 4D Gaussian primitive comes from prior work (4D-GS, 4D-Rotor; Sec. 1–2). Our contribution is a new modeling view of EEG: we are the first to cast EEG super-resolution as Gaussian splatting, with explicit Gaussian sources in brain space "rendered" to electrode positions. Prior EEG-SR methods operate entirely in electrode space, and NBF uses an implicit per-sample MLP without explicit sources. Making this work for EEG required domain-specific design, each validated by ablation:

  • Sources anchored in brain space. Free-floating centers migrate to the scalp and become shortcut interpolators (App. F.3); a cortical-shell grid raises SEED NMSE by +116% (Table 2).
  • Electrodes as the "camera": rendering at 3D electrode coordinates instead of pixels (Sec. 3.2).
  • Measurement-driven sources. The LD recording drives per-grid-point amplitudes (Sec. 3.4); a global modulation instead raises Localize-MI NMSE five-fold (0.105 → 0.541).

W2 — Comparison with NBF, masked autoencoders and ZUNA.

A: We compare against seven EEG-SR methods, including the SOTA SRGDiff. The question behind NBF and masked autoencoders, whether an explicit source model helps over a black-box LD→HD mapping, is tested directly by our capacity-matched electrode-space MLP (Table 2, App. F.5): removing the source space doubles NMSE on Localize-MI (0.105 → 0.217).

We additionally evaluated ZUNA (public ZUNA1.1 weights, zero-shot, split seed 0) with the same LD electrodes, ground truth and metrics, giving unobserved electrodes as missing channels at their 3D positions (its native upsampling mode). NMSE ↓ / PCC ↑:

DatasetMethod2×4×8×16×
Localize-MIZUNA0.416 / 0.7580.467 / 0.7260.522 / 0.6930.626 / 0.635
Localize-MIEMAG0.087 / 0.9550.105 / 0.9460.140 / 0.9270.214 / 0.886
SEEDZUNA2.019 / 0.5954.342 / 0.4429.737 / 0.302–
SEEDEMAG0.156 / 0.9190.280 / 0.8480.424 / 0.758–
SEED-IVZUNA1.684 / 0.5723.079 / 0.43418.728 / 0.258–
SEED-IVEMAG0.119 / 0.9390.208 / 0.8890.340 / 0.812–

EMAG outperforms ZUNA everywhere. ZUNA's amplitude on missing channels is poorly calibrated, driving NMSE above 1 on the 62-channel caps. Even scored in its own 256 Hz band on Localize-MI, ZUNA reaches 0.125 / 0.210 / 0.317 / 0.493. The comparison is indicative: ZUNA runs at a fixed 256 Hz (Localize-MI is 8 kHz), needs ≥0.5 s inputs (Localize-MI epochs are 0.26 s), and is a 380M-parameter model versus EMAG's ~73k.


W3 — Evaluation beyond the current tasks (sleep, motor imagery, clinical).

A: Following standard practice, we evaluate on the benchmarks most commonly reported in EEG super-resolution. They span two very different regimes, focal evoked responses on a 256-channel montage (Localize-MI) and continuous affective recordings on a 62-channel cap (SEED, SEED-IV), and the ablations show the model adapting to each (Sec. 4.4). On SEED, the reconstruction also improves downstream emotion classification for all 15 subjects (App. C). Sleep, motor-imagery and clinical cohorts are valuable directions beyond the scope of this work (Sec. 6).


W4 — Does reconstruction help downstream compared with using the LD input?

A: We agree that downstream evaluation is a valuable complement to reconstruction fidelity; solving a downstream task well, however, is a substantial project in its own right. Following prior EEG-SR work, including SRGDiff, we apply a fixed, simple classifier under the same protocol to the reconstruction and to the sparse input: EMAG improves accuracy by +0.20 / +0.21 / +0.17 at 2× / 4× / 8×, for all 15 of 15 subjects (paired t-test, p < 10⁻⁸; App. C).


We hope the new ZUNA comparison and these clarifications address the reviewer's concerns, and we would be grateful if the reviewer would consider raising the score.

Thank you for the clarifications and the additional experiment results. However, some of my concerns remain unsolved. The capacity-matched MLP ablation provides only indirect evidence and does not replace a direct comparison with NBF. And no additional datasets were evaluated to address the concern about the limited range of EEG tasks. Overall, I appreciate the authors' efforts and the improvements made, but my main concerns about above and novelty remain. I will maintain my original rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.