AI may not work like the human brain after all, scientists discover

Science Daily
Read full story· 7 min
Researchers at York University found that artificial neural networks (ANNs), commonly used to model human visual processing, may not operate like the brain as previously thought. Using a novel reverse predictivity test, they showed that while AI can predict brain activity, brain activity does not predict AI model responses, revealing fundamental differences in visual information processing.

What happened

A team at York University conducted a study to test whether artificial neural networks (ANNs), often used to model human visual processing, actually function like the human brain. They applied a new method called reverse predictivity testing, which checks if brain activity can predict the internal responses of AI models. The researchers found that while AI models can predict brain activity, the reverse is not true: brain activity does not reliably predict AI model responses. This experiment involved analyzing responses to over 1,600 images, including both natural and synthetic pictures featuring common objects.

Key details

  • Study led by York University researchers, including Assistant Professor Kohitij Kar and Postdoctoral Fellow Sabine Muzellec.
  • Reverse predictivity testing was introduced as a new method to compare AI and brain processing.
  • The test used 1,320 natural and synthetic images of 10 familiar objects, plus 300 images in different visual styles.
  • AI models could predict recorded neuronal responses, but neuronal responses could not predict many AI model features.
  • No similar disconnect was found when comparing neuronal activity between different brains.

Why it matters

Neuroscientists and researchers who use AI models to study the brain may need to reconsider how much these models actually reflect biological processes. The discovery means that conclusions drawn from AI models about human vision or behavior could be misleading if the models' internal workings differ from the brain's. This affects the reliability of AI as a tool for understanding brain function and for research into neurological or psychological conditions.

#computational-neuroscience#visual-neuroscience#york-university

More in AI Research