Meta Open-Sources AI Model TRIBE v2: Accurately Predicts Brain Responses to Images, Sound and Text Without Actual Measurement

Meta's Fundamental AI Research Team (FAIR) recently open-sourced a new AI model TRIBE v2, which can accurately predict human brain responses to images, sound and text without actual measurement. This is expected to break the bottlenecks of long cycle and high cost in traditional neuroscience research.

The model works via multi-modal fusion: after receiving video, audio and text, it extracts features using Video-JEPA-2, Wav2Vec-Bert-2.0 and Llama3.2 respectively, then integrates them through Transformer architecture to output a high-precision brain activity map with 70,000 voxels (3D pixels). Its predictions are clearer than real fMRI (eliminating noise like heartbeat and slight head movement), more accurate than traditional linear models, and can replicate classic neuroscience experiments.

However, TRIBE v2 has limitations: it relies on delayed blood flow data, can't capture millisecond-level neural dynamics, and lacks touch and smell dimensions. Meta has open-sourced the code and weights, and will explore its applications in brain science experiment planning, brain-like AI construction and brain disease diagnosis in the future.