Meta Brain2Qwerty v2: the AI that transforms brain signals into text without implants

Meta has announced Brain2Qwerty v2, a new version of its artificial intelligence system designed to convert brain activity into written text without the need for brain implants or surgical interventions. The goal of the project is not to create a technology for mind reading intended for the general public, but rather to develop a tool capable of restoring communication to people who have lost the ability to speak.

The operation of the system is based on magnetoencephalography (MEG), a technique that detects weak magnetic signals produced by brain activity. During the experiments, volunteers type on a keyboard while wearing an MEG scanner. The artificial intelligence does not observe hand movements, but directly analyzes brain signals to predict what text the user intends to write.

The main evolution compared to the first version of Brain2Qwerty concerns the processing method. Previously, the system interpreted one character at a time, while the new version simultaneously analyzes letters, words, and entire sentences. To reconstruct the text, it uses large language models that can leverage context to fill in missing information, similar to the auto-suggestions found in smartphone keyboards.

Technical innovations of Meta's Brain2Qwerty v2

From a technical standpoint, Brain2Qwerty v2 combines several deep learning technologies, including Transformers and convolutional neural networks, supported by optimized language models that correct and complete data obtained from brain signals when they are incomplete or noisy. Meta has also employed artificial intelligence agents to enhance the entire decoding process and increase its efficiency in real-time. According to the research paper published by the company, the model was trained using about 22,000 sentences typed by nine participants, each of whom spent about ten hours wearing an MEG scanner during data collection sessions.

Current performance indicates an average 61% accuracy in word recognition. The best participant achieved 78% accuracy, and over half of the reconstructed sentences had at most one lexical error. Meta has also made both the code used for training and the collected dataset available as open source, aiming to promote further research in the field.

One of the most significant aspects of the project is the absence of brain implants. Unlike other high-performance brain-computer interfaces, such as Neuralink, which require electrodes surgically inserted in the brain, Brain2Qwerty v2 uses only an external scanner, eliminating the risks associated with invasive procedures. However, Meta clarifies that the technology is still far from commercial use.