foundational-brain · a visual explainer
An AI that watches the brain in motion and learns to predict what it does next — no labels, no diagnosis handed to it. Here's what it learned, and what it honestly didn't.
Built on 871 real brain scans · trained to forecast, tested for what it understands
01 / what a brain scan really is
An fMRI scanner doesn't take a photo. Every couple of seconds it measures how active hundreds of small brain regions are — producing a flickering movie of the whole brain in motion.
The grid below is one real person's brain, moment by moment. Each square is one region; red means more active than usual, blue means less. Watch the pattern churn — that restless, structured flicker is the raw material our model learns from.
02 / the idea
The brain has hundreds of regions, but at any instant they move together in a much simpler pattern. So the model does three things, over and over, entirely on its own — nobody ever tells it what a brain is “supposed” to look like.
Compress each whole-brain snapshot into a short list of numbers — its essential “state.”
From the recent states, guess the next one. This is where the “dynamics” are learned.
Expand a state back into a full brain snapshot, so the numbers stay meaningful.
Because it's only ever asked to predict the brain's own next moment, it needs no human labels at all — which means it can learn from enormous piles of ordinary scans. That's what makes it a foundation model.
03 / does it work?
Yes — clearly. Below, the blue line is what the brain actually did. The red line is the model's guess for the very next instant, made before it happened.
The test is whether it beats the laziest sensible rule — “assume the next moment looks like this one.” It beats that rule comfortably, and it does so for every single person in the held-out test set.
04 / how it “thinks”
Inside the model, each brain snapshot becomes a single point. As the brain evolves, that point traces a path — and predicting the brain becomes a matter of continuing the curve. This is the model's private map of “what brains tend to do next.”
05 / the honest part
A good demo is easy. The harder, more useful question is whether the model's “understanding” is actually about the brain. We tested it, and the answer is a genuine mix.
The prediction win is real, large, and survives on brains from scanners it never saw. There's genuine structure in how the brain moves, and the model captured it.
Its way of compressing a single snapshot is no better than a 100-year-old technique (PCA). The value is in the motion, not the snapshot.
For reading off autism diagnosis, age or sex, its features don't beat “functional connectivity,” the field's decades-old workhorse — except a little, for age.
The single thing its internal state predicts most sharply isn't the person — it's which machine did the scan. Useful dynamics, but partly tangled up with the hardware.
None of this is failure — it's the map of where to dig next. The clear next step follows straight from that last card: train the model to actively ignore the scanner, so what's left is the brain.
in one breath
We taught a machine to predict the brain's next moment from raw scans, with no labels. It genuinely learned how brain activity moves — beating the obvious baseline for every test subject and generalizing across scanners. But its “understanding” isn't yet a shortcut to diagnosis, and it partly learned the equipment. A good forecaster; not yet a mind-reader. And we know exactly what to try next.
self-supervised 871 scans encoder → RNN → decoder honest evaluation