A dataset of more than 50,000 hours of recorded implant use provided the basis for a new system to stabilize intracortical brain signals. Neuralink used this data to pretrain a brain-computer interface foundation model using self-supervised neural encoders to convert noisy spikes into stable representations. The resulting technology enables "teleporting cursors" that jump directly to targets and has achieved a control speed of 11.32 bits per second, which the company says outperforms most able-bodied people.
The updated decoders maintain functionality for weeks instead of days, reducing calibration requirements for some users from 10 min/day to 10 min/week. Future development goals include creating an API for the motor cortex and pooling data across different trial participants to enable zero-shot or one-shot calibration.
Key sources
- SOURCEhuggingnewshuggingnews.com