Neural foundation model

A neural foundation model is a large model pretrained (usually self-supervised) on heterogeneous neural recordings — across subjects, sessions, tasks and sometimes brain areas — with the goal of a reusable backbone that fine-tunes or few-shot-adapts to new decoding problems. Examples span intracortical spiking data (POYO, NDT2) and scalp or intracranial EEG (LaBraM, Brant and related). The promise is to amortize the enormous per-subject calibration cost that has kept BCIs bespoke.

The claim deserves scrutiny. Neural data lack the web-scale, standardized corpora that made foundation models work in language and vision; montages, electrode counts, sampling rates and label schemes vary widely; and evaluation is often on the same narrow benchmarks used for training. Whether current neural foundation models generalize to genuinely new subjects and hardware, rather than interpolating within a curated distribution, remains an open and actively contested empirical question.