Meta Releases TRIBE v2 Brain Model: Predicts Neural Responses Without fMRI Scans, Resolution Up 70x

For researchers relying on fMRI scans to study brain responses, high equipment costs and time-consuming experiments have long been pain points. Meta's Fundamental AI Research (FAIR) team recently launched the TRIBE v2 trimodal brain encoder model, offering a new solution to this issue.

Trained on over 1000 hours of fMRI data from 720 subjects, the model uses a Transformer architecture to generate universal neural representations across stimuli, tasks, and individuals, with resolution 70 times higher than similar systems and significantly improved speed and accuracy.

TRIBE v2 has zero-shot prediction capabilities, enabling it to predict responses from new individuals or unseen stimuli without retraining, supporting second-level virtual experiments. Meta has open-sourced the model weights, codebase, and demos, which are expected to accelerate brain-computer interface development and neurological disease research.