Why your AI vision model might be fundamentally wrong about seeing

Tanuj Joshi
Tanuj Joshi - Editorial Contributor
5 Min Read
Listen to article5 min
0%

TORONTO – For the last ten years, neuroscientists have treated artificial intelligence as a digital twin of the human eye. The assumption was simple: if a computer program can predict how a neuron fires when you look at a zebra, it must be processing the image the same way you do. A new study from York University shatters that assumption, revealing a hidden disconnect between machine logic and biological vision.

The research team, led by Kohitij Kar, Assistant Professor and Canada Research Chair in Visual Neuroscience, conducted a rigorous reverse test. They didn’t just ask if AI could mimic the brain. They asked if the brain could explain the AI. The answer was no.

The Broken Mirror of Machine Vision

The study, published in Nature Machine Intelligence, analyzed responses to 1,320 natural photographs and 300 stylized images featuring common objects like cars, faces, and apples. The AI models performed admirably in the standard forward test: they accurately predicted the activity of recorded primate neurons. But when the researchers flipped the script, using brain activity to predict the AI’s internal features, the correlation crumbled.

“The results were striking,” Kar notes. “While AI models can predict the neurons we recorded in the brain fairly well, the brain cannot equally predict many of the model’s internal features.”

This asymmetry is the key. When comparing neurons from one brain to another, the predictivity is mutual. The fact that this fails between AI and biology suggests the machines are solving visual tasks using entirely different internal strategies. They are getting the right answer, but for the wrong reasons.

“Ultimately, we need computational models to truly understand the underlying neural mechanisms of how we recognize objects. How do we see objects move? While it’s a very easy task that we do every day, computationally, though, it’s a very challenging problem,” says Kar.

For researchers studying human behavior, this is a critical warning. If the internal logic of the AI diverges from the brain, conclusions drawn about neurological conditions could be misleading.

Key Developments: York study: AI mimics brain

The stakes extend beyond computer science. Many clinical studies use these AI models to establish baselines for neurotypical perception. If the baseline is flawed, the diagnosis of deviations – such as those seen in autism or post-traumatic stress disorder – may be built on sand. Postdoctoral Fellow Sabine Muzellec, a key contributor to the study, emphasizes that the team has identified which specific components of the AI align with reality and which do not.

“Our findings challenge how similar current AI systems really are with the primate brain,” Muzellec says. “We show that models that were previously thought to be brain-like rely on internal components that the brain does not appear to use.”

This discovery does not mean AI is useless. It means it needs to be refined. The team has released a publicly available testing toolkit, allowing developers to diagnose and correct these structural mismatches. The goal is to build a new generation of models that not only recognize objects with impressive accuracy but also reflect the true biological mechanisms of perception.

A New Standard for Scientific Reliability

The study marks a shift in how the scientific community validates machine learning tools. Performance metrics alone are no longer sufficient. The focus must move to mechanistic fidelity. As AI becomes more integrated into medical diagnostics and cognitive research, this level of scrutiny is essential to prevent misinterpretation of human brain function. The path forward involves stripping away the non-biological features of current models to create a more honest digital mirror of the human mind.

Share This Article
Leave a Comment

Leave a Reply

Your email address will not be published. Required fields are marked *