🧠 Meta FAIR released Brain2Qwerty v2 - a decoder that translates brain signals into text without surgery.
The previous non-invasive approach achieved about 8% word-level accuracy. v2 averages 61% across nine participants, with the best result at 78%: more than half of the sentences had no more than one word error. For comparison, such accuracy was previously only possible with implanted electrodes.
Participants spent 10 hours typing sentences in an MEG scanner (a helmet, no surgery involved). The model was trained on 22,000 sentences - end-to-end deep learning (Transformers + CNN) plus a fine-tuned language model to recover meaning from noisy signals.
From the article: accuracy increases logarithmically with data volume - the gap with surgical methods is closing through scaling, without new architectures.
v1 was accepted in Nature Neuroscience. The code is open.
https://ai.meta.com/blog/brain2qwerty-brain-ai-human-communication/
https://facebookresearch.github.io/brain2qwerty/
#meta@rvnikita_blog #bci@rvnikita_blog #ai@rvnikita_blog #neuroscience@rvnikita_blog #research@rvnikita_blog