🧠York University Finds AI Vision Models Match Brain Output But Not Mechanism
Your brain-model analogy may be misleading
TL;DR
A York University study tested 1,320 images across 10 subjects and found AI vision models predict neural responses well but rely on internal features the brain doesn't use. Matching brain output doesn't mean matching mechanism. That matters if you're using these models to study perception or guide clinical research.
Here's the uncomfortable finding. AI vision models that predict brain activity well may be getting there by completely different routes. A new York University study shows artificial neural networks can reproduce neuron responses while relying on internal processes the brain never touches. If you're building on these models for neuroscience, this is a real problem. Researchers have leaned on ANNs for the past decade to explain the visual system, and the assumption that good prediction equals shared mechanism now looks shaky. Teams using these models to investigate human behavior or inform clinical work need to check whether the match holds beyond the surface. Does it? Not reliably, according to this work. The researchers tested 1,320 natural photographs and realistic synthetic images of 10 familiar subjects, then added 300 images of the same subjects in different visual styles. Models predicted recorded neurons fairly well. Flip the test, though, and neuronal responses poorly predicted many internal model features. Similar models already study hearing, language, and movement, so the caution reaches well past vision.

Key Points
The study examined 1,320 natural photographs and realistic synthetic images featuring 10 familiar subjects to test how well models mirror neural responses.
Keep reading — it's free
Enter your email to keep reading — plus the best of AI & tech, daily. Free, forever.
Already a member? Sign in
Comments
Be the first to comment