Can unlabelled EEG help when movement labels are scarce?

1 What the new research changes

Neuralink reports a collection exceeding 50,000 hours of intracortical recordings and masked-channel pretraining to learn decoding representations. Its presented live results use participant-specific encoders; pooling participants has not yet improved online performance. The report describes better cursor control and reduced recalibration; we found no public download for the recordings or pretrained weights in it. These are company-reported observations, not an independently reproduced benchmark [1].

The relevant idea is learning from unlabelled signals before task training. Scalp-EEG self-supervision already has precedents such as BENDR [2]. Our continuous EEG uses squared-error reconstruction rather than a spike-count Poisson objective; this tiny convolutional model does not reproduce Neuralink's architecture or scale.

2 Design and data separation

All pretraining uses only the 63 original source-training participants from PhysioNet EEGMMIDB [3]: 2,832 three-second epochs, or approximately 2.36 hours of selected windows. Existing preprocessing and the nine-channel montage are retained. Neither validation nor test signals enter self-supervised fitting.

Across three fixed seeds, both arms share the same labelled subset, supervised batch order, architecture and number of supervised epochs within each label budget. Low-label subsets are sampled within source participant and class; integer counts are reported below. The pretrained arm additionally receives 20 reconstruction epochs, so this comparison uses extra pretraining compute. Fine-tuning lasts 20 epochs, without validation-based early stopping or checkpoint selection. The smaller budget has 80 supervised updates and the full budget has 900, so the comparison across budgets does not isolate label quantity from training compute.

The nominal 10% budget takes the floor within each participant/class cell, retaining 252 labels (8.9% overall). Existing labels are consulted to construct balanced subsets, so this is a supervised-loss budget rather than a demonstrated reduction in prospective annotation costs. Masking is applied after referencing, filtering and epoch normalisation; it does not simulate physical electrode loss.

Evaluation uses 989 epochs from the 22 existing validation participants. They are excluded from fitting, but their earlier use in project development makes this exploratory evidence on a known cohort. The 21 original test participants are not evaluated. No outcome selects a new deployed model or command threshold.

The primary metric is balanced accuracy averaged equally across participants and then across seeds. Paired differences are averaged over seeds within each participant before participants are resampled for a 95% bootstrap interval. This interval is conditional on the cohort and seeds; it does not capture all uncertainty from repeated model development.

3 Recorded validation results

Brier scores describe raw probability error, with lower values better; like balanced accuracy, they are averaged equally across participants and then across seeds. No temperature calibration or abstention policy is fitted in this study. Reconstruction loss alone cannot establish useful movement classification, and these scores should not be ranked against the original test-set results from a different architecture.

Archived outputs: results JSON, predictions and split manifest, and protocol written before the run. Tables are generated from saved outputs and checked in CI.

4 Which recordings can actually be reused?

A separate paper cited by Neuralink, NoMAD, has a human intracortical dataset deposited on Dryad [4][5]. Its public metadata specifies a CC0 licence and two NWB files, but the tested download endpoints rejected access during this review. The recordings have not been ingested or used for training here.

The data-access audit preserves publisher metadata and expected checksums, with a utility to verify files if access becomes available. Published descriptions identify a different recording modality and cursor task; those data must remain separate from scalp EEG. Auditing availability is not a reproduction of NoMAD, and recorded cursor motion should not automatically be treated as an unconfounded label of intention.

5 From fast inference to useful interaction

A prospective single-centre study reported EEG-driven tone control in 13 of 41 clinically unresponsive patients after acute brain injury, with a median signal delay of 49.8 seconds in that group [6]. This task-level measure differs from model execution time; associations with recovery do not demonstrate that the BCI caused recovery.

IntentLab now separates server processing from browser round-trip time and states the three-second recorded input. Server processing includes prediction and channel-occlusion analysis, while browser round-trip includes the request and response. Neither measures live acquisition or time from intention to action. Offline filtering also uses a completed window, so low processing latency does not establish a causal streaming decoder.

A repeatable latency benchmark reports request distributions and failures. Before physical control, priorities remain independent EEG evaluation, causal processing, and unintended-command measurement during suitably defined idle periods.

6 Interpretation and limits

This experiment has one dataset, one known participant split, three training seeds and a small architecture. It does not test long-term drift, implanted electrodes, unprompted intentions or real-time headset control. Results are conditional on the label budget and training procedure; a null or negative result would not rule out self-supervised EEG learning with other data or methods. A fresh dataset and a frozen evaluation protocol are needed before broader claims.

Successful decoding does not measure consciousness. The Nature article discussed in the newspaper is a commentary about evidence for experience, not a new decoding dataset or an experimental consciousness detector [7]. IntentLab's endpoints remain observable model behaviour and stated movement labels.