Neuralink trains on 50,000 hours of brain data, hits 11.32-bit/s cursor record

Neuralink says its brain-computer interface has achieved a cursor-control record of 11.32 bits per second after training a foundation model on more than 50,000 hours of neural activity collected from its clinical trial participants.

The Elon Musk-founded brain-computer interface company used the large dataset to pretrain self-supervised neural encoders designed to interpret electrical signals recorded directly from the brain. Neuralink says the approach can make those signals more stable and useful for controlling computers.

The company reported that some decoders built using the approach continued working for weeks rather than days. For some participants, the amount of calibration required fell from about 10 minutes each day to 10 minutes per week.

Neuralink also demonstrated what it calls “teleporting cursors,” where a cursor jumps directly to an intended target rather than moving continuously across the screen. The company says its 11.32-bit-per-second result exceeds the performance of most able-bodied people, although the comparison has not been independently verified.

Brain signals keep shifting

Brain-computer interfaces translate patterns of neural activity into commands that can operate computers and other devices. But signals recorded from neurons can change over time, creating a major problem for systems intended to work reliably every day.

Neuralink’s implants record spikes, or electrical activity produced by neurons, from the motor cortex. Small changes in recorded signals can make a decoder trained on earlier data less accurate, requiring users to repeatedly recalibrate the system.

To tackle that problem, Neuralink trained what it describes as a foundation model for brain-computer interfaces. Instead of relying only on labeled examples tied to specific movements or commands, its self-supervised neural encoders learn representations from large amounts of neural activity.

The company says the resulting representations remain useful even as individual neural signals drift. That could allow a decoder to continue interpreting a user’s intended movements without requiring frequent retraining.

The dataset behind the model comes from Neuralink’s clinical trial participants, who have collectively used their implantsfor more than 50,000 hours. The company describes it as one of the largest datasets of intracortical neural activity assembled to date.

Less calibration, faster control

Reducing calibration could be particularly important for practical brain-computer interfaces. A system that needs daily adjustment places an additional burden on users, while a decoder capable of remaining stable for weeks could make an implant behave more like an everyday computer interface.

The 11.32-bit-per-second result provides another measure of that progress. Bits per second is commonly used in BCIresearch to quantify how quickly and accurately a user can select intended targets, rather than simply measuring how fast a cursor physically moves.

Neuralink is now looking at pooling neural data across multiple participants to improve its models further. Future work could also include one-shot or zero-shot calibration, potentially allowing a system to work for a new task or user with little or no additional calibration.

The company has also outlined the possibility of creating an “API for the motor cortex,” suggesting that common neuralrepresentations could eventually provide a standardized layer for translating intended movement into digital commands.

The results were released by Neuralink as a technical update and have not been presented as a peer-reviewed journal study.