Neuralink pretrains on 50,000 hours
A summary of Neuralink's update on self-supervised neural encoders.
A BCI needs a decoder, a model that turns neural activity into what the user meant to do. Every user trains their own by calibrating, which means running through structured tasks that produce labeled data. The catch is that neural signals drift. Decoders get worse within days, so users calibrate again and again. Users spent an average of 55 minutes a week just keeping control.
Meanwhile the participants had streamed over 50,000 hours of unlabeled neural data, almost none of it used. So Neuralink pretrained a per-participant encoder on it. The encoder turns noisy, drifting spikes into stable embeddings, and a small decoder is trained on top. The model is Mamba, a good fit for neural dynamics and fast at inference. Spikes go in as tokens, with each channel getting its own embedding, as in POYO. The training objective hides half the channels and has the model predict the other half.
The results:
- Six participants set personal bests. One hit 11.32 bits per second, a BCI record.
- Decoders last much longer. Some held up for three weeks, and one was still usable after 20 months. For some users, calibration dropped from 10 minutes a day to 10 minutes a week.
- It is more data-efficient. 30 seconds of labeled embeddings did about as well as 3.5 minutes of raw spikes.
- The embeddings carry more than velocity. Simple linear models can read the cursor and target position off them, which allows "teleporting" the cursor straight to the predicted target. That crossed 10 bits per second.
The open problem is pooling. Every live result used a model pretrained on one person, and models trained on several people do no better yet. There is a hint it can work, though. A model trained only on one participant transferred to another with 99% of its weights frozen, and it beat that participant's existing decoder. The long-term goals are a single calibration, then none at all, and one shared decoder for all hand movement: an API for the motor cortex.