Meta's AI research division has bridged the gap between neuroscience and machine learning, successfully decoding brain activity into typed text without surgical implants.
The Fundamental AI Research (FAIR) lab developed Brain2Qwerty, a hybrid deep-learning architecture tested in collaboration with Spain’s Basque Center on Cognition, Brain and Language (BCBL). The system interprets neural signals associated with typing.
Using magnetoencephalography (MEG), a sensor-heavy helmet measuring magnetic fields, the system achieved a 70-80% character accuracy rate. When using a standard, portable electroencephalography (EEG) cap, accuracy dropped to roughly 50%. This distinction matters: MEG machines are room-sized and costly, while EEG is widely accessible, offering a clearer path to potential real-world application.
The technology specifically decodes motor planning signals-the brain’s instructions for finger movements-rather than abstract inner thoughts. Participants in the study typed memorized sentences.
This non-invasive success is significant for populations unable to speak due to conditions like ALS or stroke. It circumvents the surgical risks associated with implants. While not a commercial product, the peer-reviewed research marks a significant milestone in Meta's long-term, open-research commitment to decoding brain signals.