/ EN
2026-09-11 00:00 Papers Foundations & Methods Translated from EN

Pretraining Cuts Labeled Data Needed for BCI Decoding by Over 90%

Summary Training a decoder to read brain signals usually means collecting a large labeled dataset from every new subject, which is slow and a burden on patients. The proposed method, MAPA, first runs self-supervised pretraining on unlabeled intracranial EEG recordings pooled across subjects, then transfers to new ones. The difficulty is that electrode contact placement and neuroanatomy vary from person to person, so MAPA adds two spatial encodings, an anatomical region embedding and a relative positional encoding, to a standard masked autoencoder. In cross-subject tests, about 164 labeled trials were enough to reach the accuracy that otherwise takes 3,500. The team reports that MAPA set new best results on the Neuroprobe benchmark in all three settings, within-session, cross-session and cross-subject, without fine-tuning, suggesting that calibration for implanted BCIs could become much shorter.
Why it matters A major obstacle to deploying implanted BCIs is that every patient must supply a large new labeled dataset to calibrate the decoder, which is slow and burdensome. This preprint brings self-supervised pretraining to intracranial EEG, using anatomical region embeddings and relative positional encoding to cope with individual differences in electrode placement and neuroanatomy, and cuts the labeled trials needed for a new subject from 3,500 to about 164. Caveat: it has not yet been peer reviewed, and benchmark scores are not clinical performance.

BCIwiki (bciwiki.com) — A self-supervised pretraining method called MAPA lets an intracranial EEG decoder reach, in a new subject, the accuracy that otherwise requires 3,500 labeled trials using only about 164. The work is by Ben Tang, Zachary Spalding and Gregory B. Cogan and was posted to the arXiv preprint server on September 11, 2026; it has not been peer reviewed.

Brain-computer interfaces decode neural activity to restore lost function, but training a high-performance decoder typically requires collecting a large labeled dataset from every new subject. Self-supervised pretraining is one way to cut that labeling cost: it learns general neural representations from unlabeled recordings accumulated across subjects. For intracranial electroencephalography, however, self-supervised learning has been difficult because contact placement and neuroanatomy differ between subjects.

MAPA adds two spatial encodings to an otherwise vanilla masked autoencoder: an anatomical region embedding and a relative positional encoding. Together, the authors say, these let the model learn neural representations that transfer to unseen subjects and across tasks. MAPA sets a new state of the art across all three regimes of the Neuroprobe benchmark — within-session, cross-session and cross-subject — without fine-tuning.

In the cross-subject regime, a linear probe on MAPA's features needs only about 164 labeled trials to reach the accuracy that takes 3,500 without pretraining, a reduction of more than 90% in labeling demand. The authors conclude that self-supervised pretraining can scale across heterogeneous intracranial EEG recordings and reduce the labeled data needed for accurate decoding in new subjects. The paper runs 15 pages including references and an appendix, with 5 figures; the code is available on GitHub.

Compiled by BCIwiki from public sources

Sources · 1
arxiv.org 2026-09-11
Read original ↗
Suggest a correction Revisions · none / EN
© 2026 BCIwiki.com Digest Topics Tips Subscribe Revisions About