SPERA Trained on 80,000 Hours of EEG From 29,048 People and Tested on 9 Tasks

The authors of the arXiv:2610.10571 preprint have introduced SPERA (Spherical Prior EEG Representation Architecture), a foundation model for electroencephalography. Pretraining used roughly 80,000 hours of recordings from 29,048 subjects across 106 datasets. On nine downstream tasks (clinical, cognitive, and brain-computer interface), the model posted the highest mean balanced accuracy among the compared solutions. The key architectural choice: SPERA predicts latent representations of the signal, not the signal itself.

Where existing EEG models break down

EEG records the brain's electrical activity cheaply and non-invasively, but building general-purpose models on it is hard. Recordings differ across subjects, hardware, and montages: some devices have 19 electrodes, others more than a hundred, and they are placed differently. A model used to one layout performs worse on another.

The second problem is the objective function. Most current EEG foundation models learn by reconstruction: mask part of the signal and ask the model to restore it. But an observed recording contains more than neural activity. It also holds non-neural components, meaning noise and artifacts. A model that tries to reproduce the signal exactly has to spend capacity on them as well.

How SPERA works

The authors use JEPA (joint-embedding predictive architecture), where prediction happens in latent space and part of the noise can be ignored. They add three EEG-specific adaptations to this scheme.

• A spatial prior based on Legendre polynomials. It is built into attention and encodes the geometry of electrode placement on the scalp. This lets a single model work with different montages.
• Factorized attention. Temporal and spatial attention alternate, with full-attention blocks inserted at regular intervals.
• A relational spectral regularizer. It aligns the similarity structure in latent space with spectral representations of the signal, that is, with the frequency rhythms on which a large part of EEG analysis is built.

Key parameters from the abstract:

• pretraining volume: about 80,000 hours of EEG;
• subjects: 29,048;
• data sources: 106 datasets;
• evaluation: nine downstream tasks, balanced accuracy metric;
• the authors also claim parameter efficiency under linear probing and robustness to different recording conditions.

The last point matters in practice. Linear probing freezes the backbone and trains only a linear head on top of the representations. If quality holds up, the embeddings themselves are informative and do not require heavy fine-tuning for each task.

Where this could be useful and what to watch

For now this is a preprint, and the abstract gives no specific metric values or names of the models SPERA was compared against. The claim of leading on mean accuracy needs to be checked against the full tables: an average over nine tasks can hide failures on individual ones. It is also worth checking which datasets went into pretraining and whether they overlap with the test sets, a common weak point in EEG benchmarks.

If the claims hold up, three groups benefit. Clinical teams can build detectors on small labeled samples by training a linear head on top of a ready-made backbone. BCI developers may depend less on a specific headset and electrode layout. Researchers working on cognitive tasks get a single base model instead of a set of narrow architectures.

The practical test for any team is this: take your own data with a nonstandard montage, freeze SPERA, train a linear classifier, and compare it with your current pipeline. If the gain holds, the approach of latent prediction with a geometric prior will become established as a standard path for EEG.