Meta releases TRIBE v2 for 700-volunteer brain activity prediction
Meta AI has introduced TRIBE v2, a new AI foundation model designed to predict how the human brain processes complex visual and auditory stimuli. This model, trained on extensive fMRI data from 700 volunteers watching varied media, aims to accelerate neuroscience research by enabling brain activity predictions with 70x higher resolution than previous models, without requiring human subjects for every experiment. Meta AI is openly releasing the model, codebase, paper, and an interactive demo to foster further research and applications in neuroscience and AI system development.
Key Takeaways
- TRIBE v2 is Meta AI’s first model for human brain responses to sights, sounds, and language.
- Training used fMRI recordings from more than 700 healthy volunteers exposed to images, podcasts, videos, and text.
- Meta says TRIBE v2 can make zero-shot predictions for new subjects, languages, and tasks.
- The company says the model consistently outperforms standard modeling approaches and predicts high-resolution fMRI brain activity.
- Meta is releasing the model weights, codebase, paper, and an interactive demo under a CC BY-NC license.
Why It Matters
Meta is packaging neuroscience as a foundation-model problem: TRIBE v2 predicts high-resolution fMRI activity without needing human subjects for every experiment. That matters immediately for researchers trying to test hypotheses faster, and it could also feed back into AI system development, since Meta says brain insights can guide AI from neuroscientific principles. The open release of the model, code, paper, and demo lowers the barrier for follow-on work outside Meta. What to watch next is whether researchers use the demo and open weights to reproduce the 70x resolution claim on new subjects, languages, and tasks.
Read full article at ai.meta.com