Neuralink says decoders pretrained on 50,000 hours of brain data cut recalibration and set a cursor record
Self-supervised models trained on participants' own unlabeled recordings kept cursor decoders usable for weeks, and one participant reached 11.32 bits per second, the company reports.
On October 1, Neuralink reported that it trained participant-specific neural encoders on more than 50,000 hours of unlabeled recordings streamed by its trial participants over two years. The encoders turn drifting spike data into steadier representations for cursor decoding. The company says six participants set personal bests, three passed its earlier record of 10.39 bits per second, and one reached 11.32 bits per second. Some decoders stayed usable for more than three weeks, and some users’ calibration time fell from about 10 minutes a day to 10 minutes a week.
Frequent recalibration is a major burden for home BCI users. These are company-reported results in a blog post, not peer-reviewed, and Neuralink says pooling data across participants has not yet improved live performance.
Learn more: Information transfer rate · Neuralink