The system records brain activity using a helmet-like magnetoencephalography (MEG) scanner, a non-invasive brain imaging device commonly used in neuroscience research. It then feeds those raw neural signals into an end-to-end AI model that reconstructs the sentences a person is trying to type. Meta said it further improves accuracy by fine-tuning large language models on neural data, allowing the system to use semantic context when interpreting noisy brain recordings.
“We trained Brain2Qwerty v2 on approximately 22,000 sentences from nine volunteer participants, each recorded for 10 hours wearing a magnetoencephalography (MEG) device while actively typing,” Meta wrote. “Instead of relying on hand-crafted pipelines to detect neural events, we use end-to-end deep learning to decode directly from raw brain signals.”
Meta said Brain2Qwerty achieved a 61% average word accuracy, compared with roughly 8% for previous non-invasive methods. The company is releasing the system's code and dataset as part of its Digital Brain Project, which also includes a $5 million fund to support open neuroscience datasets.
Meta said Brain2Qwerty v2 approaches levels of accuracy previously achieved only with techniques requiring brain surgery. The company said its non-invasive approach could help bridge the gap between invasive neuroprosthetics and communication systems that do not require surgery.
“Our hope is that this work, done in the open, advances neuroscience to identify, diagnose, and treat neurological disorders faster than in siloes,” Meta wrote.
Meta did not immediately respond to a request for comment by Decrypt.
