Product1 publisher3 min readPublished
Weizmann's AI decoder redraws what eight scanned volunteers were looking at
Weizmann researchers built an AI decoder that redraws images people saw from brain scans, trained on eight volunteers who each viewed about 9,000 pictures. The uses for patients are still hopes, and one neuroscientist is already warning about consent.
The Product Desk · Product desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- The decoder runs two branches, one predicting an image's layout and colors and one its content, and their output steers a diffusion model that draws the picture.
- Short of scan data, the team trained a second model that predicts brain activity from an image, then used the two models together to improve each other.
- The scans came from newer public datasets recorded on higher-resolution scanners, with each voxel covering about one cubic millimeter of brain.
- Michal Irani hopes the tool will show more about how the brain works and could perhaps help locked-in people communicate or let scientists recreate dreams.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint Any use outside a lab inherits the decoder's inputs: a person who sat through thousands of images on a scanner finer than the standard machine.
- decision Labs reusing public brain-scan datasets now have to decide whether the consent those volunteers originally gave covers training a decoder of what they saw.
- exposure If a similar decoder ever works without the subject's cooperation, people's mental imagery becomes readable without their consent, the scenario Sprague called worrisome.
A volunteer lies inside a giant magnet that tracks oxygenated blood moving through the brain [6]. Pictures go past one after another, about 9,000 of them [2]. Eight people did this. Multiply it out and the decoder rests on roughly 72,000 image viewings [1], all recorded by other research groups and released publicly [5].
Today the tool is for those volunteers and the neuroscientists who study them. The pitch is about patients who cannot speak. What the team actually built is narrower: it rebuilds pictures that cooperative people in a research scanner were looking at, with what Technology Review called remarkable precision [1]. Earlier attempts in the field produced blurry images that were hard to make sense of [17]. Irani's complaint about current rival tools is subtler. If a person saw a banana, those models could generate a banana, but it would look different. "It wouldn't have the same structure, the same position," she said [18][8].
The scanners were also finer than typical ones. A typical fMRI voxel covers about three cubic millimeters and around 16,000 neurons [7]. The datasets Irani used resolve voxels a third that size [4]. If neurons are spread evenly, each of those voxels holds roughly 5,300 [2].
The training loop is the team's answer to the data shortage [10]. Give the encoder a new picture, a leopard in Irani's example, and it predicts the scan a person would produce. The decoder then tries to rebuild the leopard from that predicted scan. Irani said the first rebuild probably won't look much like a leopard, but repeating the cycle eventually leads to dramatic improvements [11]. New training pairs can come from pictures alone instead of from new scanner sessions.
Outside reaction divides along the two possible uses. Judy Illes, a neuroethicist at the University of British Columbia who was not involved, called the work "magnificent" [12]. She said using the approach therapeutically, to help people with neurologic conditions, is tremendously exciting [13]. Tommy Sprague, a neuroscientist at the University of California Santa Barbara, said "The results seem very impressive" [14]. He then added: "But if there's a way to surreptitiously extract information about what you're thinking about, then...150 years of sci-fi can come true anytime, and that's worrisome in a lot of ways" [15].
Technology Review's account does not say whether the decoder works on anyone outside the eight people it learned from. The clinical case and Sprague's worry both depend on that answer.
For an ethics board, or a lab deciding whether to build on this work, a two-by-two sorts the claims. One axis asks whether the person being decoded agreed to it and supplied their own training scans. The other asks whether the setup needs a high-resolution research scanner. I'd treat this paper as a research instrument and require every follow-on claim to say which cell it sits in. The tradeoff is that promising therapeutic work will be asked for evidence it may not have yet. Illes's patients belong in the consent row. Sprague's scenario, mental imagery read without consent, is the other row [16]. Every result described so far comes from one cell: volunteers who agreed, scanned on one-millimeter hardware [2][4].
What to watch
- Results from the same approach on standard fMRI scanners with voxels near three cubic millimeters, not just one-millimeter research machines.
- Any test with a locked-in patient, the first clinical use Irani names.
- Whether ethics guidance on consent for reusing public brain-scan data follows the warnings from Sprague and others.