Overview
- Meta has unveiled Brain2Qwerty v2, a groundbreaking AI system capable of translating brain activity into text without surgical procedures.
- This new model boasts an average word accuracy of 61%, a significant improvement over the approximate 8% accuracy of earlier non-invasive techniques.
- Meta has made the training code for both Brain2Qwerty v1 and v2 available, with its research partner providing the dataset for v1.
On Monday, Meta announced the release of Brain2Qwerty v2, an innovative AI system that translates brain activity into written text through non-invasive brain recordings. This research aims to assist individuals who have lost their ability to communicate due to brain lesions.
The technology utilizes a helmet-like magnetoencephalography (MEG) scanner, a non-invasive device widely used in neuroscience. The system captures raw neural signals and employs an end-to-end AI model to reconstruct the sentences a person intends to type. Meta enhances accuracy by fine-tuning large language models based on neural data, which helps the system interpret noisy brain recordings with contextual understanding.
Meta disclosed, “We trained Brain2Qwerty v2 on around 22,000 sentences from nine volunteers, each recorded for 10 hours while using a magnetoencephalography (MEG) device during typing.” This approach shifts from traditional methods that rely on manually crafted pipelines to an end-to-end deep learning strategy that decodes directly from raw brain signals.
The Brain2Qwerty system has achieved an impressive 61% average word accuracy, in stark contrast to the roughly 8% accuracy seen in prior non-invasive methods. As part of its Digital Brain Project, Meta is sharing the code and dataset, alongside establishing a $5 million fund aimed at supporting open neuroscience datasets.
Meta noted that as the volume of training data increased, decoding accuracy also improved, indicating that further data could enhance performance. The company mentioned that AI agents tested various optimizations for the decoding process before finalizing the training configuration.
In a related study published in Nature Neuroscience, Meta's researchers pointed out that despite the significant advancements in brain-to-text decoding through AI, most effective brain-computer interfaces still rely on surgically implanted electrodes, which pose scalability challenges due to the associated surgical risks and maintenance issues over time.
Meta claims that Brain2Qwerty v2 reaches accuracy levels comparable to those achieved only through surgical methods. The firm believes this non-invasive approach could help close the gap between invasive neuroprosthetics and communication systems that do not require surgery.
“We hope that this open-source work will accelerate neuroscience research to identify, diagnose, and treat neurological disorders more swiftly than in isolated environments,” Meta stated.
This announcement coincides with a surge in brain-computer interface research, which includes projects by Elon Musk’s Neuralink and Merge Labs, backed by OpenAI CEO Sam Altman, aiming to restore communication for those with neurological impairments.
While firms like Neuralink and Synchron are focused on surgical interfaces, an increasing number of researchers and startups are leveraging AI to enhance non-invasive systems. In September 2024, startup Neurable launched AI-driven EEG headphones aimed at monitoring cognitive focus and fatigue. A year later, MIT spinout AlterEgo introduced a device that translates silent neuromuscular signals from the face and throat into text and commands, offering a practical alternative to implanted brain-computer interfaces.
Meta did not provide an immediate response to a request for comment from Decrypt.