This research was made possible by our great team
@lucyzmf.bsky.social, Svetlana Pinet, Jérémy Rapin, Hubert Banville, Stéphane d'Ascoli and
@jeanremiking.bsky.social, AI at Meta, @bcbl.bsky.social, @cnrs.fr, @psl-univ.bsky.social, and Hôpital Fondation Rothschild!

Interested in how the brain actually transforms the representation of a sentence into a sequence of motor action? Go check our companion Neuroscience paper:
🧵6/7

While not yet practical for everyday use, this could one day pave the way to restore communication for patients who can’t speak or move – without requiring invasive implants.
🧵5/7

Brain2Qwerty consists of 3 modules: convolutions, transformer and character-level language model.
It achieves ~20% character-error-rate (CER) on the best individuals. That’s a huge leap from EEG-based methods (67% CER)!
🧵4/7

Here, we develop a non-invasive method to decode text from brain activity.
We recorded 35 volunteers while they typed sentences in the MEG scanner and trained Brain2Qwerty to decode their keystrokes.
🧵3/7

The past five years have been marked by a rapid progress in brain-computer-interfaces.
Electrodes implanted inside the motor cortex can now be input to an AI system to restore communication: www.nature.com/articles/s41...
However, this approach requires neurosurgery.
🧵2/7