Researchers have released Brain2Qwerty v2, a non-invasive AI pipeline that decodes real-time sentences from magnetoencephalography (MEG) recordings without surgical implants. The system achieves a 61% word accuracy rate overall and up to 78% for top performers, significantly outperforming previous non-invasive methods.
- Trained on approximately 22,000 sentences from nine participants wearing MEG devices while typing.
- Utilizes end-to-end deep learning to decode directly from raw brain signals rather than hand-crafted pipelines.
- Fine-tunes large language models on neural data to leverage semantic context and bridge noisy inputs.
- Deploys AI agents to explore decoding pipeline optimizations, with final configurations selected by engineers.
- Releasing full training code for v1 and v2, alongside the v1 dataset from partner BCBL.
This research aims to provide a scalable alternative to invasive neuroprosthetics for millions of people with brain lesions that prevent communication. By advancing open foundational models of the brain, the authors hope to accelerate the identification and treatment of neurological disorders.