AI Could be Using Dangerous Shortcuts to Solve Complex Recognition Tasks
Researchers found that deep convolutional networks are insensitive to object configural properties.
Deep convolutional networks (DCNNs), which are based on configural shape perception, do not perceive things the same as humans (which could be detrimental in real-world AI). According to Professor James Elder who is the co-author of a York University paper published recently in iScience, DCNNs do not perceive things like humans (through configural shape perception). This could be harmful for real-world AI applications.
The study was conducted by Elder who is the York Research Chair for Human and Computer Vision and Co-Director at York’s Centre for AI & Society. Nicholas Baker, a psychology assistant professor at Loyola College, Chicago, and a former VISTA Postdoctoral Fellow at York, found that deep learning models failed to capture the configural aspect of human shape perception.