Brain Foundation Models Divide Neuroscientists on What Counts as Understanding

Meta has committed $15 million to build a brain foundation model from 15,000 hours of precision-scanning fMRI data, while Stanford, Neuralink and other teams pour tens of millions more into the same wager: that training artificial neural networks on vast brain recordings will yield models that keep improving. Whether that bet pays off divides the field, and a Dual Perspectives article published today in The Journal of Neuroscience stages the argument directly.

Bing Wen Brunton, professor of biology at the University of Washington, and Patrick Mineault, NeuroAI research lead at the Amaranth Foundation, debate whether AI models that predict brain activity can also explain how the brain works, and whether that understanding requires modeling the body. They will revisit the topic during a live session at the Society for Neuroscience annual meeting in Washington, D.C., on Sunday, 15 November.

What can these models already do?

The concrete results are accumulating. A foundation model of the mouse visual cortex, trained on calcium imaging (a technique using fluorescent indicators to monitor neuronal firing) of mice watching natural videos, predicted neuronal responses to coherent motion and noise, according to a Nature study published last year. The model also predicted neurons' anatomical cell types, dendritic features and connections, and could model the activity of new mice after seeing less than half an hour of their brain data.

Meta's TRIBE v2, drawing on more than 1,000 hours of brain scans from people watching videos, listening to audio and reading text, predicts fMRI responses to new stimuli in people it was never trained on, according to a preprint posted on arXiv in May. Meta declined to comment for this story.

Andreas Tolias, professor of ophthalmology at Stanford and co-leader of the Amaranth-funded Enigma Project, worked on the Nature study. His lab's earlier model predicted that when a mouse's pupil dilates, its visual neurons become more sensitive to ultraviolet light — a prediction the team confirmed through in-vivo experiments. "These models can also hallucinate," Tolias says; a model earns trust only when it can "make a novel discovery that is verifiable in the brain."

Does more data always help?

The scaling story is mixed. OmniMouse, trained on recordings of more than 2.3 million neurons from 73 mice, improved with more data, but adding parameters helped only up to a point. A 2025 study found decoding models improved with more recording hours per person, yet adding more people barely helped. Three foundation models, including BrainLM, trained on 6,700 hours of fMRI, predicted cognitive scores worse than a simple linear model of brain connectivity, and a larger version performed worse than a smaller one, per a May arXiv preprint.

"I'm skeptical that foundation models will be just as successful for every type of data," Brunton says. "We need to understand scaling laws of models, for each type of data, to understand how much data is enough." Grace Lindsay of New York University, unaffiliated with those models, counters that current performance is not the ceiling: "More data should help."

Prediction versus explanation

Mineault argues data-driven models outperform task-trained alternatives. The best task-trained vision models explain only about half of the variance in neuronal recordings from the inferotemporal cortex, per the Brain-Score benchmark. Training directly on recordings from intermediate visual areas such as V1, V2 and V4 could constrain models to become more brain-like, he says.

Daniel Yamins of Stanford, involved in the fMRI- and MEG-based Stanford Digital Brain Project, notes that better prediction does not equal brain-like internals. "They'll just look like a prediction machine," he says; combining task-trained and data-driven approaches is ideal, but how to do so "hasn't been figured out." John Krakauer of Johns Hopkins calls a perfect mimic "a second brain you have to neurally understand" and is "increasingly skeptical" of claims extending to drugs or lesions.

Brunton's thought experiment frames her caution: a transformer trained on thousands of filmed falling apples would predict the next fall "better than F=ma even could," she writes, but "you could not use this transformer model to design a rocket and put it on the Moon." Both authors converge on similar next steps — measure as much as possible, then distill principles — with Mineault starting from large-scale recordings and connectomes, and Brunton welcoming digital twins only if they are embodied. "We have to admit the possibility that the answer could be, 'Oh yeah, this is not going to work,'" she says.