The University of York found that even the smartest AI models could not match human visual processing

Professor James Elder, co-author of a study published at the University of York, said that deep convolutional neural networks (DCNNs) do not use configured shape perception to observe objects the way humans do, which can be dangerous in real-world AI applications

Deep Learning Models Fail to Capture the Configurative Nature of Human Shape Perception is a collaborative study conducted by Elder and Nicholas Baker, an assistant professor of psychology at Loyola College in Chicago, published in the journal iScience in Cell Press

The study used a new type of visual stimulus known as "Frankenstein" to explore how the human brain and DCNNs handle holistic, configured object properties

"Frankenstein is just an object

The researchers found that the human visual system was confused by Frankenstein, while dcnn didn't — revealing insensitivity

"Our findings explain why deep AI models fail under certain conditions and point to the need to consider tasks beyond object recognition in order to understand visual processing in the brain," Elder said

One such application is the traffic video security system: "In a busy traffic scene, objects such as vehicles, bicycles and pedestrians obstruct each other, and in the driver's eyes, they are a pile of irrelevant debris," Elder explains

According to the researchers, the modifications to training and architecture made to make the network more like a brain did not result in configuration processing, and none of the networks were able to accurately predict human object judgments

York University is a modern, multi-campus urban university located in Toronto, Ontario