Meta Unveils AI for Decoding Brain Activity

On June 29, Meta unveiled a groundbreaking system capable of reconstructing typed text from brain activity without the need for surgical procedures. The new technology, Brain2Qwerty v2, achieved an average word-level accuracy of 61%, a significant improvement over the 8% accuracy seen in prior non-invasive methods.

Brain2Qwerty v2 was developed using data from around 22,000 sentences and tested on nine healthy volunteers. Each participant underwent 10 hours of scanning in a magnetoencephalography machine while typing phrases they heard.

“Instead of manually created schemes to identify neural events, we use end-to-end deep learning to decode data directly from raw brain signals,” the company explained.

The system is designed to work with ongoing brain activity, enabling it to reconstruct full sentences. It utilizes a refined language model that incorporates semantic context to enhance accuracy. The highest-performing participant achieved a word-level accuracy of 78%, with over half of their decoded sentences containing one or no errors.

Despite these advancements, Brain2Qwerty v2 is not yet ready for clinical or everyday applications. Its current testing involved healthy individuals who were actively typing, and further validation is required for those with speech or mobility impairments.

Moreover, the technology does not provide real-time text input, as it processes entire sentences rather than displaying each word immediately after typing. This project is part of the Digital Brain Project, which aims to establish open neuroscience datasets and has a funding allocation of $5 million.

In September 2022, Meta also released research on a "brain decoder" that leverages AI to translate thoughts into speech.