Tri-modal encoding
Pretrained embeddings capture features shared by AI models and the human brain — V-JEPA2 for video, Wav2Vec-BERT for audio, and LLaMA 3.2 for language.
Stage 01 · V-JEPA2 · Wav2Vec-BERT · LLaMA 3.2
In-silico neuroscience
Predicting how the brain responds to sight, sound and language.
TRIBe v2 is Meta's foundation model that simulates fMRI brain activity in response to naturalistic stimuli — a digital mirror of human neural processing, open for research exploration.
Explore examples
See the brain respond,
example by example.
Select a stimulus to see the Neuroanalytical model's predicted brain activity beside it. These are precomputed examples — the model maps each input onto the cortical surface.
Want to run your own media? The full interactive demo, built by Meta, is available on Meta's original demo.
Why this matters to Smart Bridge
Where digital strategy meets neuroscience.
Smart Bridge includes professionals with backgrounds in psychology and neuroscience. This page showcases the Neuroanalytical model as an open research tool — exploring how humans process complex stimuli, aligned with our research-informed approach to audience behaviour.
- In-silico experiment simulation
- Multimodal stimulus analysis
- Behavioural science research
- Non-commercial research exploration
What the Neuroanalytical model does
Brain response prediction,
not general image analysis.
The Neuroanalytical model maps multimodal stimuli onto the cortical surface and predicts how an average human brain would respond — enabling in-silico experiments without new fMRI recordings.
Architecture
Three stages to brain prediction.
The Neuroanalytical model combines state-of-the-art encoders with a unified Transformer that maps multimodal representations onto individual fMRI voxels across the whole brain. It predicts activity across 70,000 voxels — trained on more than 700 volunteers — and consistently outperforms standard modelling approaches.
Universal integration
A Transformer learns representations shared across all stimuli, tasks, and individuals — enabling zero-shot generalisation to new subjects and media types.
Stage 02 · Unified Transformer
Brain mapping
A subject layer maps universal representations onto fMRI voxels — 3D pixels tracking neural activity through slow changes in blood flow and oxygenation.
Stage 03 · ~70,000 voxels · fsaverage5
The Neuroanalytical model is developed by Meta AI and released under the CC BY-NC 4.0 licence for non-commercial research. Model weights, code, and demo are provided by Meta — Smart Bridge hosts this page as a research showcase only. Sample video clip: Big Buck Bunny © Blender Foundation, CC BY 3.0. The interactive 3D brain mesh is derived from Brain for Blender by Anderson M. Winkler, licensed CC BY-SA 3.0. Brain activity graphics are illustrative placeholders.