acceptodds
Under review as a conference paper at ICLR 2027

vNeuro: Foundation for EEG Vision Models

Abstract

EEG foundation models are typically pretrained from scratch on large corpora of neural recordings. Yet experienced clinical neurophysiologists interpret EEG largely by visual inspection, recognizing spikes, rhythmic bursts, and time–frequency patterns by eye. This raises the question of whether representations learned from natural visual data can be reused for EEG. We introduce vNeuro, a framework that adapts a frozen, video-pretrained V-JEPA 2.1 encoder to EEG through a lightweight trainable interface. vNeuro converts multichannel time–frequency representations into visual tokens that preserve electrode geometry, and trains this interface on unlabeled EEG through latent forecasting with distribution regularization, while the visual backbone remains fixed. During downstream training, only the adapter and classification head are optimized. With a ViT-B backbone, vNeuro achieves balanced accuracies of 81.40% on TUAB abnormality detection and 68.35% on TUEV event classification (68.73% with ViT-G), competitive with EEG foundation models pretrained on tens of thousands of hours of EEG. Ablation and representation analyses show that both the EEG-to-visual token mapping and the regularized self-supervised adaptation contribute to performance. These results suggest that pretrained visual encoders can serve as effective EEG backbones without pretraining a large EEG-specific model.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG representation learning, visual structural priors, time-frequency modeling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_X65QOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes vNeuro, an EEG foundation modeling framework that transfers pretrained visual representations to EEG analysis. Specifically, the authors develop an EEG-to-visual adapter that transforms multichannel time–frequency representations into visual-compatible tokens while incorporating electrode topology. A frozen V-JEPA 2.1 encoder is then adapted through latent temporal forecasting and distribution regularization. Experiments on TUAB and TUEV demonstrate competitive performance against existing EEG foundation models.

Critical strengths
  1. The idea of leveraging pretrained visual representations for EEG analysis is interesting and provides an alternative to conventional EEG-specific foundation model pretraining.

  2. The proposed EEG-to-visual adapter incorporates both time–frequency characteristics and electrode topology, providing a reasonable interface between EEG signals and visual encoders.

  3. The experimental results are promising, and the authors provide additional ablation studies and representation analyses to investigate the effectiveness of the proposed framework.

Critical weaknesses
  1. Although the authors compare pretrained and randomly initialized visual encoders in Appendix C.4, this analysis is limited to DINOv3. A controlled comparison using the main V-JEPA backbone would more directly establish whether its visual pretraining contributes to the reported improvements.

  2. Experiments are conducted only on TUAB and TUEV, which originate from the same EEG corpus. Cross-dataset evaluation, particularly across different electrode configurations and recording environments, would better substantiate the claimed foundation modeling capability.

  3. The TUEV Cohen's kappa and weighted-F1 values for the EEG-CWT adapter differ between Tables 1 and 2. These discrepancies should be carefully checked and corrected to ensure the reliability of the reported results.

Questions for authors
  1. The framework predicts the representation of a future EEG window from preceding windows. However, the authors do not sufficiently explain why this objective is particularly suitable for learning discriminative EEG representations. Further analysis is needed to establish the relationship between temporal predictability and downstream classification performance.

  2. The adapter initially mixes information from different EEG channels through convolution and compresses the signals into a relatively small visual token grid. This process may discard fine-grained temporal information or channel-specific characteristics. The author should investigate whether the clinically relevant electroencephalogram patterns are preserved during this transformation process.

  3. In Equations (2) and (4), the authors directly introduce several numerical values, such as 0.09. What is the rationale behind these values, and how were they determined?

  4. EEG signals are highly susceptible to artifacts and noise, which may affect the quality of time–frequency representations. The authors should investigate the robustness of the proposed framework under different noise levels to assess its reliability in practical EEG analysis.

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

vNeuro turns 16 bipolar EEG leads into per-lead CWT scalograms and passes them through a trainable adapter (local CNN, electrode-geometry gating, temporal mixing) into a frozen V-JEPA 2.1 encoder. The adapter is first trained without labels by forecasting the latent state of a fifth frame from four context frames, with SIGReg to prevent collapse, then fine-tuned with a ResNet18 and MLP head. The paper reports 81.40% balanced accuracy on TUAB and 68.35% on TUEV (68.73% with ViT-G) and claims gains over EEG foundation models.

Critical strengths
  1. The question is worth asking, and the appendix gives a clean partial answer. A frozen pretrained DINOv3-B on CWT maps reaches 67.5% balanced accuracy on TUEV while the same frozen network with random weights sits at chance, and 17 input transforms are compared under one probe. That transfer result is the most informative part of the paper.
Critical weaknesses
  1. Table 1 does not match the cited papers, although the text says baselines are taken from the original publications. LaBraM reports 81.40% (Base) to 82.58% (Huge) on TUAB and 64.09% (Base) on TUEV, but Table 1 lists 77.35% and 55.50%. CBraMod sits at exactly 50.00% on TUAB, which is chance, while the CSBrain paper lists it at 78.91% on TUAB and 66.71% on TUEV. Against published numbers vNeuro-B ties LaBraM-Base on TUAB balanced accuracy with lower AUROC, trails it on TUEV kappa and weighted F1, and the 11.77-point TUEV gain becomes about 2.
  2. The proposed parts are not shown to matter. The frozen DINOv3-B probe, with no adapter and no forecasting, gets 67.5% on TUEV against 68.35% for vNeuro-B, within one standard deviation. The tables also disagree: TUEV kappa and weighted F1 for the same adapter are 0.7198 and 0.8510 in Table 2 but 0.6498 and 0.8110 in Tables 1, 3 and 4, with identical deviations. If Table 1 is right, the 3D CNN V-JEPA interface in Table 2 beats vNeuro on both.
Questions for authors
  1. Clarification: where do the LaBraM, CBraMod and REVE rows in Table 1 come from, and why is CBraMod at chance on TUAB?
  2. Clarification: which table is correct for TUEV kappa and weighted F1?
  3. Experiment: run the frozen DINOv3 and V-JEPA probes under the exact Table 1 protocol and splits.
  4. Clarification: the TUAB seeds repeat only Stage 2 from one Stage 1 checkpoint. Please report seeds over the full pipeline.
Reviewer_FFosOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes vNeuro, a framework that adapts a frozen, video-pretrained V-JEPA encoder for EEG classification. It converts EEG into time–frequency representations and learns an EEG-to-vision adapter through latent temporal forecasting with SIGReg regularization, followed by supervised training of the adapter and classification head. Experiments on TUAB and TUEV report best balanced accuracies of 81.40% and 68.73%, respectively, with additional analyses of input representations, regularization, and positional conditioning.

Critical strengths

A promising approach of applying visual representations to EEG. The paper claims that a frozen video-pretrained encoder can support EEG classification through a trainable EEG interface. Results on TUAB and TUEV demonstrate that this approach offers a way to reuse existing visual backbones without EEG-specific pretraining of the backbone.

Critical weaknesses

The core methodological contribution is insufficiently validated. The ablation studies do not sufficiently establish the benefits of the proposed visual prior and temporal learning objective. Table 3 compares forecasting with and without SIGReg, but the absence of a supervised-only baseline leaves the contribution of the forecasting stage unclear. Similarly, the pretrained-versus-random comparison is conducted in a separate DINOv3 probing setup, leaving the effect of video pretraining in the actual vNeuro pipeline unverified. Without these controls, I am not convinced that the reported gains can be attributed to the mechanisms presented as the paper’s core contributions.

The foundation-model claim goes beyond the evidence. The evaluation uses only two closely related clinical EEG datasets, TUAB and TUEV. The authors do not show that one EEG-adapted checkpoint transfers across different tasks or recording setups. The 17 input representations are also tested with a separate probing protocol. This says little about how robust the full vNeuro system is. Overall, the experiments do not support the claim of a broadly reusable EEG foundation model.

The paper requires substantial improvement in writing and organization, particularly in Section 3. Key terms such as “lead,” “frame,” “context,” “target,” and “cross-window” are insufficiently explained when introduced, making the spatial and temporal structure difficult to follow. Figure 2 also lacks a clear explanation of how its panels connect and what the “shared mapping” denotes. The research questions are poorly connected to the corresponding discussions, while phrases such as “17 signal-processing” are imprecise. Although the appendix provides additional details, the main text needs clearer definitions, more consistent terminology, and a more coherent presentation of the method.

Questions for authors

The paper motivates forecasting as a means of learning EEG-specific temporal dynamics. Beyond the analysis of SIGReg, what evidence shows that forecasting captures temporal relationships that improve downstream EEG representations? Since the majority foundation model idea use random masked among tokens to grasp representation, why does forecasting idea work?

Can the authors isolate the benefit of video pretraining within the actual vNeuro pipeline through matched comparisons with other image-pretrained/ video pretrianed alternative? The separate probing experiments in Appendix C do not directly establish this benefit. For instance, how about sam_v3 or sdxl? How about doing sft over the pretrained encoder?

Is Stage 1 trained separately for TUAB and TUEV, or is a shared EEG-adapted checkpoint transferred between datasets? What data are used for this stage, and what transfer evidence supports the claimed generality of the resulting EEG representation? How about applying this model to other EEG dataset that is well explored in EEG foundation model, say sleeping data, emotion, stress and motor imaginary?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.