We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Hacker News·Jun 30·5 min read
Meta's brain-scanning system reads sentences non-invasively, code open source
TL;DR
Many individuals with brain lesions struggle to communicate due to traditional invasive methods being difficult to scale. Brain2Qwerty v2 is a non-invasive AI tool that decodes brain activity into coherent sentences in real-time.
✦ Why It Matters
Engineers and researchers can leverage non-invasive brain decoding techniques to develop accessible communication aids for individuals with speech impairments.
Key Takeaways
- Explore non-invasive methods for brain signal decoding.
- Utilize end-to-end deep learning for real-time applications.
- Consider fine-tuning language models on neural data.
- Investigate potential applications in assistive communication technologies.
Full Summary
Individuals with brain lesions often face communication challenges, and existing invasive techniques like stereotactic electroencephalography are not easily scalable. Brain2Qwerty v2 is a cutting-edge AI model that decodes brain activity into text without requiring surgical implants.
It was trained on 22,000 sentences from nine participants using magnetoencephalography (MEG), a non-invasive brain imaging technique. Instead of traditional methods, it employs end-to-end deep learning to interpret raw brain signals directly.
The model leverages fine-tuned large language models to enhance semantic understanding, resulting in coherent sentence recovery from noisy inputs. Brain2Qwerty v2 achieved a word accuracy rate of 61%, with the best participant reaching 78%, marking a significant improvement over previous non-invasive approaches.
This advancement has the potential to greatly enhance communication for those with severe speech impairments.