Cajal-1: 218M-Parameter Model Predicts Whole-Cortex fMRI Responses from Media

·2026.10.19 09:54

Key point

The 218-million-parameter model predicts fMRI responses at 20,484 cortical points every 1.49 seconds without a scanner.

Cajal-1 is a 218-million-parameter model trained on fMRI data from 2,000 people watching films, TV, and images.
On films held out by Meta for testing TRIBE v2, Cajal-1 is 25% more accurate and closer to real brain activity in 27 of 32 recordings.
The model predicts fMRI responses at 20,484 points across the human cortex every 1.49 seconds for any video, podcast, or song.
The model and a Colab demo are available on Hugging Face.

Named after Santiago Ramón y Cajal, the first Spaniard to win a Nobel Prize in science.
Operates at the same speed as standard processing but requires no physical scanner.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.