Your Vision Foundation Model Is Secretly an EEG Foundation Model

Guoan Wang, Feng Liu

November 2026

TL;DR SeeEEG draws EEG as images that a frozen DINOv3 vision transformer, never trained on EEG, can read, and learns only the display (montage, filters, contrast; at most 1,113 parameters per view). It beats seven frozen EEG foundation models in balanced accuracy on all eight datasets, with a mean rank of 1.3 against 2.9 for the best baseline.

What they do

EEG foundation models are pretrained on thousands of hours of recordings, yet their frozen features transfer poorly to new tasks. SeeEEG instead uses a frozen image model, DINOv3 ViT-B/16, that has never seen EEG. It learns only how the EEG is drawn for the backbone, plus a light readout head. The authors compare it with seven frozen EEG foundation models on eight datasets: event types, abnormal EEG, sleep staging, dementia, slowing, emotion, imagined speech and motor imagery.

How it works

Token-local rendering. Each channel gets its own row of 16-pixel patches, so one token holds one channel over 80 ms. The channel is drawn either as grayscale amplitude at a fixed gain or as a 1-64 Hz wavelet scalogram that spans the patch height.
Learned display. Montage weights (referential, common average, Laplacian and bipolar are special cases), a display filter and contrast settings start from the clinical montage chosen on validation. They are trained with the task loss through the frozen backbone, using 29 to 1,113 parameters per view.
Cross-fitting. Each training window updates the display only through a readout head that never saw its subject, so the display can’t learn subject-specific traits.

Results

With DINOv3, SeeEEG beats every EEG foundation model in balanced accuracy on all 8 datasets. Its mean rank is 1.3 in kappa/AUROC, against 2.9 for REVE. Kappa margins reach +27.1 on slowing and +8.1 on imagined speech.
The same displays fed to V-JEPA 2.1 beat every EEG foundation model on both metrics on 6 of 8 datasets.
Mean balanced accuracy rises from 50.5 for a plain page of traces to 64.6 for the full pipeline. The best EEG foundation model scores 60.0. On motor imagery, the learned display adds 10.2 points over the best clinical montage.
A randomly initialized ViT-B scores 15.2 points lower on average.

Limitations

All results are offline classification of fixed windows from public datasets. The paper tests no online BCI use and reports no inference latency. The display variant, readout head and backbone block are chosen per dataset on validation, so each new task needs its own fitting.

Figures

From Figure 2: Overview of SeeEEG. (a) The learned display sets the montage, display filter and contrast, starting fr... — From Figure 2: Overview of SeeEEG. (a) The learned display sets the montage, display filter and contrast, starting from a clinical montage. (b) Each channel fills its own row of 16-pixel tokens as a waveform (value view) or a scalogram, and long windows are wrapped into a near-square image. (c) A frozen pretrained ViT extracts the class token and mean patch token of each block. (d) A linear probe or token MLP reads these tokens.

From Figure 3: Learned displays for one motor-imagery (PhysioMI) and one emotion (SEED) window: value strips of three... — From Figure 3: Learned displays for one motor-imagery (PhysioMI) and one emotion (SEED) window: value strips of three channels before and after optimization, the change in each channel's neighbor weights, and the validation kappa of a linear probe on each channel's frozen tokens.

From Figure 4: Evidence maps on TUEV. Top: the most confidently recognized test window of each event type, with token... — From Figure 4: Evidence maps on TUEV. Top: the most confidently recognized test window of each event type, with token-level evidence overlaid on the scalograms of the five bipolar pairs with most evidence. Bottom: share of evidence on each pair, averaged over the recognized test windows of each class.

Abstract

Electroencephalography (EEG) foundation models are pretrained on thousands of hours of recordings, yet their frozen features transfer poorly to downstream tasks, and this transfer does not consistently improve with larger models or more pretraining data. We show that a vision foundation model that has never seen EEG is a stronger frozen EEG encoder than these specialized models, once the signals are rendered as images that a vision transformer can read. Our method, SeeEEG, keeps the vision backbone frozen and learns only how EEG is displayed to it. Each channel occupies a dedicated row of patches and is drawn either as a waveform at a fixed gain or as a scalogram whose frequency range spans the patch height, so that task-relevant patterns fall within individual tokens. The montage, filtering, and contrast parameters of this display are initialized from a standard clinical montage and optimized with the task loss through the frozen backbone, which amounts to at most 1,113 trainable parameters per view. We compare SeeEEG with seven EEG foundation models on eight clinical, sleep, affective, and brain-computer interface datasets, with every encoder frozen and paired with identical readout heads. With a frozen DINOv3 ViT-B, SeeEEG outperforms every EEG foundation model in balanced accuracy on all eight datasets and reaches a mean rank of 1.3 in Cohen kappa or AUROC, compared with 2.9 for the strongest baseline, REVE. The learned displays transfer without further optimization to the video encoder V-JEPA 2.1, which outperforms every EEG foundation model on both metrics on six of the eight datasets. Replacing DINOv3 with a randomly initialized ViT-B lowers the score by 15.2 points on average, so the ability to read rendered EEG comes from visual pretraining.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong