York Study Finds Brain-Like Vision AI Uses Parts the Brain Doesn’t
A York University test found that AI vision models predict brain cells well, but brain activity predicts many of their internal features poorly.

Artificial neural networks that closely predict activity in the brain’s visual system may be solving the same tasks with very different internal machinery. Researchers at York University in Canada reported the result in Nature Machine Intelligence, according to a ScienceDaily summary published Nov. 21.
For years, neuroscientists have used these networks, computer systems trained to recognize objects in images, as stand-ins for the brain’s visual system. A model earned its place by predicting how recorded neurons respond to pictures. The York team asked whether the relationship also holds in the other direction.
How the team ran the test in reverse
The group, led by Assistant Professor Kohitij Kar, built a method it calls reverse predictivity testing. Instead of asking how well a model predicts neurons, it asks how well the recorded neurons predict what happens inside the model. If a network really works like the brain, the researchers reasoned, prediction should work both ways.
The images included 1,320 natural photographs and realistic synthetic pictures of 10 subjects: a bear, an elephant, a face, an apple, a car, a dog, a chair, a plane, a bird and a zebra. The subjects appeared against a variety of natural scenes and indoor or outdoor backgrounds. The team added 300 more images showing the same subjects as outlines, drawings, simplified shapes and artistic interpretations.
What the comparison showed
The models were fairly good at predicting the neurons the team recorded. The reverse was not true. Neuronal responses did a much poorer job of predicting many of the models’ internal features.
The gap did not appear when the team compared neurons from one brain with neurons from another. “The brain cannot equally predict many of the model’s internal features,” Kar said.
In Kar’s reading, today’s AI systems solve visual tasks partly with internal strategies the brain may not use. He also said the parts of the models that do line up with the brain are better at predicting real human behavior. Sabine Muzellec, a postdoctoral fellow on the team, said models long considered brain-like rely on components the brain does not appear to use.
Which models can still be trusted for brain research
Kar warned that if researchers keep refining models only by how well they predict neural activity, the differences from real primate brains could build up over time. That would weaken the main reason neuroscientists use the models, which is to get realistic explanations of how the brain works.
The team does not argue for dropping the networks. Its approach is meant to pick out which parts of a model truly match brain activity, so researchers can build more reliable ones. Kar said this matters for his group’s autism research, which uses models of the neurotypical brain (a brain with typical patterns of development and functioning) as a baseline.
Similar networks are already used to study hearing, language and movement. The researchers say those systems may also depend on internal processes that differ from biological ones, though this study tested vision only.
How sure is this?
The work is published in a peer-reviewed journal, Nature Machine Intelligence, and the summary was provided by York University. The sources give the image counts above but do not say how many neurons or how many AI models were tested, so the size of the effect across the field is not clear from this report.
The finding concerns visual object recognition. Whether networks used for hearing, language or movement show the same mismatch has not been tested in this work.
The team has released a public testing toolkit so other AI developers can run the same check on their own models. Kar said the aim is to develop systems that more faithfully reproduce how the brain makes sense of what it sees, which could eventually help research on conditions such as post-traumatic stress disorder and autism.
Source: ScienceDaily
Site editor: Nikita Sheikin



