pith. sign in

arxiv: 2606.23707 · v1 · pith:BEAUNHCOnew · submitted 2026-06-12 · 📡 eess.SP · cs.AI· cs.LG

Coordinate-Queryable Neural Field Reconstruction for EEG Spatial Super-Resolution with Unseen-Electrode Generation

Hongjun Liu , Leyu Zhou , Zijianghao Yang , Chao Yao This is my paper

Pith reviewed 2026-06-27 05:12 UTC · model grok-4.3

classification 📡 eess.SP cs.AIcs.LG
keywords EEG spatial super-resolutionunseen electrode generationconditional neural fieldimplicit neural representationmissing channel reconstructioncoordinate queryscalp field modeling
0
0 comments X

The pith

A coordinate-queryable neural field learns a continuous scalp field from variable observed EEG channels to reconstruct signals at any electrode position, including those never seen during training.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that EEG spatial super-resolution should be treated as learning one shared conditional scalp field instead of fixed mappings between preset electrode sets. A position-guided encoder turns any subset of observed channels and their coordinates into a latent condition, while a conditional implicit neural representation decoder produces the signal value at any queried coordinate. This design directly supports random missing channels and strict unseen-electrode generation because the field is defined everywhere once the condition is set. Experiments on multiple datasets show clear gains, including a 37.5 percent NMSE reduction and 2.12 dB SNR improvement under held-out electrode conditions. If the reformulation holds, EEG super-resolution becomes robust to the changing channel patterns that occur in real deployments.

Core claim

The central claim is that EEGSR reduces to learning a shared conditional scalp field from partially observed support channels: a position-guided encoder summarizes the available EEG values and coordinates into a latent condition, and a conditional implicit neural representation decoder reconstructs target signals by querying that condition at desired electrode coordinates; a fidelity-preserving channel corruption strategy during training further stabilizes the encoded representation so that the resulting field remains consistent with the observed data.

What carries the argument

The conditional implicit neural representation decoder that reconstructs EEG signal values at arbitrary scalp coordinates conditioned on a latent encoding of observed channels and positions.

If this is right

  • The same trained model can reconstruct signals at electrode locations never present in the training data.
  • Performance improves under both random missing-channel patterns and strict held-out electrode settings across multiple EEG datasets.
  • The fidelity-preserving corruption strategy produces a more stable latent condition that better constrains the decoded field.
  • Inference requires only the current support channels and their coordinates plus the query coordinates, with no retraining for new layouts.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The approach could support online adaptation when electrodes are added or removed during a recording session without retraining.
  • It might extend naturally to source localization tasks that require field values at continuous scalp locations rather than discrete electrodes.
  • Similar coordinate-queryable conditioning could be applied to other spatially sampled biosignals such as ECG or EMG grids.

Load-bearing premise

EEG signals everywhere on the scalp are samples from one continuous field whose value at any coordinate depends only on a latent condition derived from whatever subset of channels is observed.

What would settle it

A controlled test in which the model is given support channels from one spatial distribution but asked to reconstruct a held-out electrode whose true value deviates sharply from any continuous field consistent with the support set, such as an isolated artifact at that location.

Figures

Figures reproduced from arXiv: 2606.23707 by Chao Yao, Hongjun Liu, Leyu Zhou, Zijianghao Yang.

Figure 1
Figure 1. Figure 1: Overview of ScalpINR. Partially observed EEG support channels are encoded into a support-conditioned scalp field, which is queried by a conditional INR decoder to reconstruct missing channels or generate unseen-electrode signals. index. The decoder predicts the target EEG value by querying a coordinate-conditioned neural field: xˆ(pq, t) = fω(z, ϕ(pq), τ (t)), (4) where ϕ(pq) is the embedded query coordina… view at source ↗
Figure 2
Figure 2. Figure 2: Detailed architecture of ScalpINR. The encoder aggregates position-guided channel tokens into a latent condition, and the coordinate-queryable INR decoder predicts EEG signals at target electrode coordinates. electrodes improve robustness to unstable channel quality, and visible-but-predicted electrodes prevent the model from treating observed channels as unconstrained pass-through signals. Together, these… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative waveform comparison under random missing-channel reconstruction. Rows correspond to different support ratios r, and curves show the ground-truth signal and reconstructions produced by ESTformer, MCMA, and ScalpINR. ScalpINR better follows the temporal morphology of the ground truth, especially under lower support ratios where fixed-layout baselines exhibit stronger amplitude attenuation or loca… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative waveform comparison under strict unseen-electrode generation. The target electrodes are permanently held out from the support set during training and are queried only during evaluation. Rows correspond to different support ratios r, and columns show representative examples from AAD, BCI2000, and SEED. Compared with spherical spline interpolation and ZUNA, ScalpINR produces waveforms that are mo… view at source ↗
read the original abstract

EEG spatial super-resolution (EEGSR) in real deployments is challenged by random channel missingness, unstable electrode quality, and changing visible-channel patterns caused by bad contacts or device variability. Most existing EEGSR methods learn a fixed low-to-high channel mapping under pre-defined input-output layouts, which makes them brittle when missing channels vary at test time. In this paper, we reformulate EEGSR as learning a shared conditional scalp field from partially observed support channels. Specifically, a position-guided encoder summarizes the observed EEG channels and their coordinates into a latent condition, and a conditional implicit neural representation decoder reconstructs target EEG signals by querying this condition at desired electrode coordinates. During inference, the model directly reconstructs unseen electrode signals from the available EEG support and the queried coordinates. To strengthen the constraint of the encoded latent representation on the decoder and thereby construct a more stable scalp field consistent with the observed channels, we further introduce a fidelity-preserving channel corruption training strategy under mixed electrode states. Extensive experiments across multiple EEG datasets demonstrate the effectiveness of our framework for both random missing-channel reconstruction and strict unseen-electrode signal generation. Notably, under the strict held-out-electrode setting on AAD, our method reduces NMSE by 37.5\% and improves SNR by 2.12 dB over the strongest baseline, showing its ability to synthesize signals at electrode locations never exposed during training.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper reformulates EEG spatial super-resolution (EEGSR) as learning a shared conditional scalp field from partially observed support channels. A position-guided encoder produces a latent condition from observed EEG channels and their coordinates; a conditional implicit neural representation (INR) decoder then reconstructs signals at arbitrary target coordinates. A fidelity-preserving channel corruption strategy is used during training to stabilize the latent representation. Experiments across multiple datasets report gains for both random missing-channel reconstruction and strict unseen-electrode generation, with a 37.5% NMSE reduction and 2.12 dB SNR improvement over the strongest baseline on the AAD dataset under held-out electrodes.

Significance. If the empirical results are reproducible, the coordinate-queryable formulation could meaningfully improve robustness to variable electrode layouts and missing channels in real-world EEG deployments. The adaptation of conditional INRs to produce a continuous scalp field from sparse, position-aware observations is a coherent modeling choice that directly targets the brittleness of fixed-layout EEGSR methods.

major comments (1)
  1. [Experiments] Experiments section: the central claim of 37.5% NMSE reduction and 2.12 dB SNR gain under strict held-out-electrode setting on AAD rests on comparisons whose protocol (data splits, baseline implementations, statistical tests, and exact electrode hold-out procedure) is not provided in sufficient detail to allow independent verification. This is load-bearing for the empirical contribution.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their positive assessment of the coordinate-queryable INR formulation and for highlighting the need for greater experimental transparency. We address the single major comment below.

read point-by-point responses

  1. Referee: [Experiments] Experiments section: the central claim of 37.5% NMSE reduction and 2.12 dB SNR gain under strict held-out-electrode setting on AAD rests on comparisons whose protocol (data splits, baseline implementations, statistical tests, and exact electrode hold-out procedure) is not provided in sufficient detail to allow independent verification. This is load-bearing for the empirical contribution.

    Authors: We agree that the reported gains on the AAD held-out-electrode task require a fully specified protocol to support independent verification. In the revised manuscript we will expand the Experiments section with: (i) the exact train/validation/test split ratios and subject-wise partitioning for AAD; (ii) complete implementation details and hyper-parameter settings for every baseline, including any modifications required to operate under the strict unseen-electrode protocol; (iii) the statistical tests (including p-values and correction method) used to establish significance of the 37.5 % NMSE and 2.12 dB SNR improvements; and (iv) the precise electrode hold-out procedure, including how support-channel sets are sampled, how target coordinates are defined, and how the fidelity-preserving corruption schedule is applied during training. These additions will be placed in a dedicated subsection and will be accompanied by a supplementary table listing all numerical settings. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical result on external benchmarks

full rationale

The paper reformulates EEGSR as learning a conditional scalp field via a position-guided encoder and INR decoder, then reports empirical gains (37.5% NMSE reduction, 2.12 dB SNR) under strict held-out-electrode evaluation on AAD and other datasets. No equations, fitted parameters, or self-citations are shown that reduce any claimed prediction or uniqueness result to the model's own inputs by construction. The architecture and corruption strategy are modeling choices whose validity is tested externally rather than defined circularly. This is a standard non-circular empirical modeling paper.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim depends on the domain assumption that scalp EEG forms a continuous, queryable field and on the architectural choice of a latent condition that generalizes across electrode subsets; both are introduced by the paper rather than inherited from prior literature with independent verification.

free parameters (1)
  • neural network weights and hyperparameters
    The encoder and decoder parameters are fitted to EEG training data to realize the conditional field representation.
axioms (1)
  • domain assumption EEG signals on the scalp can be modeled as samples from a continuous function that remains consistent across different subsets of electrode positions
    Invoked when reformulating EEGSR as learning a shared conditional scalp field from partially observed support channels.
invented entities (1)
  • latent condition produced by the position-guided encoder no independent evidence
    purpose: To compress observed EEG channels and coordinates into a representation that the decoder can query at arbitrary target coordinates
    New architectural component introduced to enable coordinate-based reconstruction; no independent falsifiable evidence supplied outside the model itself.

pith-pipeline@v0.9.1-grok · 5791 in / 1363 out tokens · 22697 ms · 2026-06-27T05:12:31.335133+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 4 canonical work pages

  1. [1]

    Multi-channel masked autoencoder and comprehensive evaluations for reconstructing 12-lead ecg from arbitrary single-lead ecg.NPJ Cardiovascular Health, 1(1):34, 2024

    Jiarong Chen, Wanqing Wu, Tong Liu, and Shenda Hong. Multi-channel masked autoencoder and comprehensive evaluations for reconstructing 12-lead ecg from arbitrary single-lead ecg.NPJ Cardiovascular Health, 1(1):34, 2024

    2024

  2. [2]

    Corley and Yufei Huang

    Isaac A. Corley and Yufei Huang. Deep eeg super-resolution: Upsampling eeg spatial resolution with generative 10 USTB@AI3D T able 8.Strict unseen-electrode generation results under different support ratiosr, evaluated only on permanently held-out electrodes. Here, r denotes the fraction of non-held-out electrodes used as visible support, andr = 1means tha...

    work page doi:10.1109/bhi.2018.8333379 2018

  3. [3]

    Spectral features of resting-state EEG in parkinson’s disease: A multicenter study using functional data analysis.Clinical Neurophysiology, 151:28–40, 2023

    Alberto Jaramillo-Jimenez, Diego A Tovar-Rios, Johann Alexis Ospina, Yorguin-Jose Mantilla-Ramos, Daniel Loaiza-López, Verónica Henao Isaza, Luisa María Zapata Saldarriaga, Valeria Cadavid Castro, Jazmin Ximena Suarez-Revelo, Yamile Bocanegra, et al. Spectral features of resting-state EEG in parkinson’s disease: A multicenter study using functional data a...

    2023

  4. [4]

    Neural brain fields: A nerf-inspired approach for generating nonexistent eeg electrodes.arXiv preprint arXiv:2601.00012, 2025

    Shahar Ain Kedem, Itamar Zimerman, and Eliya Nachmani. Neural brain fields: A nerf-inspired approach for generating nonexistent eeg electrodes.arXiv preprint arXiv:2601.00012, 2025

    arXiv 2025

  5. [5]

    Temporal variational implicit neural representations.arXiv preprint arXiv:2506.01544, 2025

    Batuhan Koyuncu, Rachael DeVries, Ole Winther, and Isabel Valera. Temporal variational implicit neural representations.arXiv preprint arXiv:2506.01544, 2025

    arXiv 2025

  6. [6]

    Super-resolution for improving eeg spatial resolution using deep convolutional neural network—feasibility study.Sensors, 19(23):5317, 2019

    Moonyoung Kwon, Sangjun Han, Kiwoong Kim, and Sung Chan Jun. Super-resolution for improving eeg spatial resolution using deep convolutional neural network—feasibility study.Sensors, 19(23):5317, 2019

    2019

  7. [7]

    EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces.Journal of neural engineering, 15(5):056013, 2018

    Vernon J Lawhern, Amelia J Solon, Nicholas R Waytowich, Stephen M Gordon, Chou P Hung, and Brent J Lance. EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces.Journal of neural engineering, 15(5):056013, 2018. 11 USTB@AI3D r=0.5r=0.25r=0.125 AmplitudeAmplitudeAmplitude Time Time TrueESTformerMCMAScalpINR AAD BCI2000 SEED ...

    2018

  8. [8]

    A protocol for comparing dry and wet eeg electrodes during sleep.Frontiers in neuroscience, 14:586, 2020

    Sven Leach, Ku-young Chung, Laura Tüshaus, Reto Huber, and Walter Karlen. A protocol for comparing dry and wet eeg electrodes during sleep.Frontiers in neuroscience, 14:586, 2020

    2020

  9. [9]

    Estformer: Transformer utilising spatiotemporal dependencies for electroencephalogram super-resolution.Knowledge-Based Systems, 317:113345, 2025

    Dongdong Li, Zhongliang Zeng, Zhe Wang, and Hai Yang. Estformer: Transformer utilising spatiotemporal dependencies for electroencephalogram super-resolution.Knowledge-Based Systems, 317:113345, 2025

    2025

  10. [10]

    Imputeinr: time series imputation via implicit neural representations for disease diagnosis with missing data

    Mengxuan Li, Ke Liu, Jialong Guo, Jiajun Bu, Hongwei Wang, and Haishuai Wang. Imputeinr: time series imputation via implicit neural representations for disease diagnosis with missing data. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, pages 9241–9249, 2025

    2025

  11. [11]

    Step-aware residual-guided diffusion for eeg spatial super-resolution.The Fourteenth International Conference on Learning Representations, 2026

    Hongjun Liu, Leyu Zhou, Zijianghao Yang, and Chao Yao. Step-aware residual-guided diffusion for eeg spatial super-resolution.The Fourteenth International Conference on Learning Representations, 2026

    2026

  12. [12]

    Eeg signal reconstruction using a generative adversarial network with wasserstein distance and temporal-spatial-frequency loss.Frontiers in neuroinformatics, 14:15, 2020

    Tian-jian Luo, Yachao Fan, Lifei Chen, Gongde Guo, and Changle Zhou. Eeg signal reconstruction using a generative adversarial network with wasserstein distance and temporal-spatial-frequency loss.Frontiers in neuroinformatics, 14:15, 2020

    2020

  13. [13]

    Signal quality evaluation of an in-ear eeg device in comparison to a conventional cap system.Frontiers in Neuroscience, 18:1441897, 2024

    Hanane Moumane, Jérémy Pazuelo, Mérie Nassar, Jose Yesith Juez, Mario Valderrama, and Michel Le Van Quyen. Signal quality evaluation of an in-ear eeg device in comparison to a conventional cap system.Frontiers in Neuroscience, 18:1441897, 2024

    2024

  14. [14]

    Time Series Continuous Modeling for Imputation and Forecasting with Implicit Neural Representations.Transactions on Machine Learning Research Journal, April 2024

    Etienne Le Naour, Louis Serrano, Léon Migus, Yuan Yin, Ghislain Agoua, Nicolas Baskiotis, Patrick Gallinari, and Vincent Guigue. Time Series Continuous Modeling for Imputation and Forecasting with Implicit Neural Representations.Transactions on Machine Learning Research Journal, April 2024. doi: 10.48550/arXiv.2306.05880. URLhttps://hal.science/hal-04759780

    work page doi:10.48550/arxiv.2306.05880 2024

  15. [15]

    The target electrodes are permanently held out from the support set during training and are queried only during evaluation

    Petr Nejedly, Vaclav Kremen, Vladimir Sladky, Mona Nasseri, Hari Guragain, Petr Klimes, Jan Cimbalnik, 12 USTB@AI3D r=0.5r=0.25r=0.125 AmplitudeAmplitudeAmplitude Time Time TrueSphericalSplineZUMAScalpINR AAD BCI2000 SEED Time Figure 4.Qualitative waveform comparison under strict unseen-electrode generation. The target electrodes are permanently held out ...

    2019

  16. [16]

    Csdi: Conditional score-based diffusion models for probabilistic time series imputation.Advances in neural information processing systems, 34:24804–24816, 2021

    Yusuke Tashiro, Jiaming Song, Yang Song, and Stefano Ermon. Csdi: Conditional score-based diffusion models for probabilistic time series imputation.Advances in neural information processing systems, 34:24804–24816, 2021

    2021

  17. [17]

    Generative ai enables eeg super-resolution via spatio-temporal adaptive diffusion learning.IEEE Transactions on Consumer Electronics, 71 (1):1034–1045, 2025

    Shuqiang Wang, Tong Zhou, Yanyan Shen, Ye Li, Guoheng Huang, and Yong Hu. Generative ai enables eeg super-resolution via spatio-temporal adaptive diffusion learning.IEEE Transactions on Consumer Electronics, 71 (1):1034–1045, 2025. doi: 10.1109/TCE.2025.3528438

    work page doi:10.1109/tce.2025.3528438 2025

  18. [18]

    Zuna: Flexible eeg superresolution with position-aware diffusion autoencoders.arXiv preprint arXiv:2602.18478, 2026

    Christopher Warner, Jonas Mago, JR Huml, Mohamed Osman, and Beren Millidge. Zuna: Flexible eeg superresolution with position-aware diffusion autoencoders.arXiv preprint arXiv:2602.18478, 2026

    arXiv 2026

  19. [19]

    Geometry-and relation-aware diffusion for eeg super-resolution.arXiv preprint arXiv:2602.02238, 2026

    Laura Yao, Gengwei Zhang, Moajjem Chowdhury, Yunmei Liu, and Tianlong Chen. Geometry-and relation-aware diffusion for eeg super-resolution.arXiv preprint arXiv:2602.02238, 2026

    arXiv 2026

  20. [20]

    Investigating critical frequency bands and channels for EEG-based emotion recognition with deep neural networks.IEEE Transactions on Autonomous Mental Development, 7(3):162–175,

    Wei-Long Zheng and Bao-Liang Lu. Investigating critical frequency bands and channels for EEG-based emotion recognition with deep neural networks.IEEE Transactions on Autonomous Mental Development, 7(3):162–175,

  21. [21]

    doi: 10.1109/TAMD.2015.2431497. 13

    work page doi:10.1109/tamd.2015.2431497 2015