Stanford launches an AI brain model project: about ten volunteers will each undergo a hundred MRI sessions and a hundred sessions wearing a helmet with magnetic sensors in the first phase

Stanford launches an AI brain model project: about ten volunteers will each undergo a hundred MRI sessions and a hundred sessions wearing a helmet with magnetic sensors in the first phase

On November 2, Stanford’s Wu Tsai Neurosciences Institute announced the five-year Stanford Digital Brain Project, supported by philanthropists Mark and Mary Stevens and Clara Wu Tsai. Brain recordings will be used to train a “large brain model” along the lines of language models. The team has pledged to make the data and models available to researchers worldwide.

Just as a language model is trained to predict the next word, the brain model will first learn to predict brain activity at the next moment. It should then learn to predict activity across the whole brain from the videos participants watch, their behavior, and their gaze direction, explained Russell Poldrack, one of the project’s leaders.

Daniel Yamins, one of the project’s founders, points to findings from the past ten years: neural networks trained solely on tasks such as object recognition or language processing can accurately predict activity in brain regions that perform the same tasks. Yamins compares training a neural network directly on brain recordings to moving “from theoretical science to empirical science.” Until now, a shortage of data has stood in the way. “Our main work is collecting data,” Yamins told The Transmitter.

The first of ten recording rounds is expected to last a year, with a planned start in early 2027. About ten volunteers will watch a wide variety of videos. A model exposed to only a small fraction of human experience will model only that fraction, Poldrack says. Each volunteer will undergo roughly a hundred MRI sessions and as many sessions wearing a helmet with magnetic sensors. At about an hour per session, that amounts to roughly a hundred hours per device. Between 2012 and 2014, Poldrack scanned his own brain more than a hundred times, but to his knowledge, no one has collected a comparable set of recordings from one person using both devices.

Poldrack describes the small number of participants as “both a limitation and an advantage”: scanner time is finite, and averaging across a group erases subtle features of each person’s brain that are needed for personalized models. According to Yamins, hundreds of hours of recordings per person should allow the team to build a detailed model of one brain and adapt it to other people for whom much less data is available. Meta, meanwhile, trained its brain model TRIBE v2 on MRI recordings from 25 people, averaging 18 hours per person.

An MRI scanner with a 7-tesla magnet (hospital scanners use 1,5 to 3 tesla) can locate activity to within a millimeter. It does so by tracking blood flow to active neurons, but that blood flow lags behind neural activity by seconds. The magnetoencephalography helmet uses sensors to detect magnetic fields generated by neurons firing. It resolves timing to within a millisecond but is less precise at locating activity. When the two recordings agree, Poldrack says, researchers can be much more confident that errors specific to either device have been ruled out. The project’s hypothesis goes further: a model required to match both recordings and behavior will provide more detail than either recording alone.

Yamins envisions using the model to simulate treatment. Magnetic brain stimulation is used to treat depression that does not respond to medication, and researchers are still working to identify the best stimulation settings. A digital brain could allow them to test those settings before treating a patient.

Bing Wen Brunton of the University of Washington, who is not involved in the project, asks: “What evidence is there that this will work, beyond our hope?” Caterina Gratton of the University of Illinois Urbana-Champaign sees the small group as a test of the central premise: “Can we build any kind of digital brain that reasonably simulates even one person?” If so, Gratton adds, “that already gives us a starting point.”