Meta Brain2Qwerty v2: Writing by Thought Without Surgery

Meta Brain2Qwerty v2: An Unprecedented Advancement in Translating Thoughts into Text

Meta has recently introduced Brain2Qwerty v2, an enhanced version of its innovative system that translates brain signals into text without the need for surgical implants. This technology relies on a magnetoencephalography (MEG) device, a sophisticated scanner that captures the magnetic fields generated by brain activity. Study participants use this device while typing on a keyboard, allowing the AI to directly decode brain signals to guess the text the user intends to write.

Significant Improvements Over the First Version

The first iteration of Meta Brain2Qwerty was limited to decoding letters one by one. The v2 version, on the other hand, has the capability to interpret not only characters but also entire words and phrases. This feat is made possible through the use of large language models that fill in missing information, similar to predictive text on smartphones. Meta claims that this method allows the system to grasp the overall meaning of a sentence, facilitating the reconstruction of a coherent message even when brain signals are difficult to interpret.

In the background, several deep learning technologies collaborate. Models such as Transformers and convolutional neural networks are integrated with language models that act as smart correctors. When certain information is incomplete or garbled, context helps accurately estimate the user's intent. Meta has also employed AI agents to optimize the decoding process and enhance real-time performance.

To train its model, Meta collected approximately 22,000 phrases typed by nine volunteers, each spending nearly ten hours under an MEG scanner during the training sessions. Currently, Brain2Qwerty v2 achieves an average accuracy of 61% in word recognition, with the best participant reaching a rate of 78%, and over half of the decoded phrases containing only one word error.

Towards Public Use?

The most accurate brain-computer interfaces often rely on electrodes implanted directly in the brain, such as those developed by Neuralink from Elon Musk. While this method offers high precision, it requires surgical intervention. Meta's Brain2Qwerty v2 takes a different approach by using only an external MEG scanner to analyze brain activity. This method avoids the risks associated with intracranial implants while achieving promising results.

However, Meta is still far from offering a product intended for the general public. The MEG scanners used by Brain2Qwerty v2 are bulky, expensive, and reserved for research laboratories. Therefore, it won't be possible to compose emails by thought in the near future. Despite these limitations, Meta believes that the progress made shows that non-surgical brain-computer interfaces are becoming increasingly credible.

For individuals who are speech-impaired, this technology could represent a far more significant advancement than current chatbots or image generators. Meta has also released the training code and dataset as open source, allowing other teams to continue this research.