What if AI could help scientists better understand the mysteries of how the brain allows humans to see 鈥 and be fooled by 鈥 the world around them?
鈥淭here鈥檚 a lot we don鈥檛 understand about how we see the world,鈥 says Richard Murray, a professor in 91亚色鈥檚 Faculty of Health and director of the Murray Lab, which studies visual perception.
One of those open questions is lightness perception, which involves the brain鈥檚 ability to judge how light or dark a surface really is, even when lighting conditions change. For example, it allows a white shirt to appear white in both shadow and sunlight.
Known as lightness constancy, this ability plays a central role in how people recognize objects and make sense of their surroundings.
Yet despite its importance, it has proven difficult to explain in terms of how this process happens in the brain, and how to recreate it in models researchers can study.
In recent years, advances in AI have started to change that picture. In computer vision 鈥 a field focused on teaching machines to interpret images 鈥 deep learning models have become increasingly good at solving one of the central challenges of lightness perception: separating lighting from the properties of the surfaces themselves.
That progress raised a new question for Murray. Could these systems help explain how human perception works?
鈥淚 wondered whether AI鈥檚 vision might be similar to ours, and whether it could be a good starting point for theories of human vision,鈥 he says.
The result of that question is published in , with PhD student Jaykishan Patel as first author and postdoctoral researcher Alban Flachot as second author.
To test that idea, Murray and his collaborators turned to a classic tool from vision science: visual illusions.
Among those used by researchers to understand how the brain interprets light and shadow are images in which two areas have the same actual shade, but appear lighter or darker depending on their surroundings. For example, a grey square placed in shadow can look lighter than an identical grey square in a bright area.
Murray wanted to know whether an AI model trained to interpret images would be fooled in the same way.
To find out, the team built an AI model to discern whether a surface is truly light or dark, or simply appears that way because of lighting.
The system was not trained on visual illusions; instead, researchers tested it using the same kinds of illusion images shown to people, to see if it would be fooled. The researchers wanted to see whether the AI would behave the same, or correctly identify them as identical.
鈥淚 was surprised,鈥 says Murray. 鈥淓ven though these systems aren鈥檛 trained to mimic human vision, they find many of the same things easy or difficult. They also see many of the same illusions.鈥
In the study, the AI did, in fact, 鈥渇all for鈥 those illusions. When shown images where identical patches were placed in different lighting conditions, it often judged them as different 鈥 like human observers do.
Just as important was what didn鈥檛 happen. When the same system was trained on a different task 鈥 like cleaning up blurry or distorted images 鈥 it no longer showed the same illusion effect. This means the illusion only appears when AI is trying to make sense of what it is looking at, not when it is simply processing the image.
The findings are promising, but they also raise new questions.
鈥淲e鈥檝e shown that artificial vision is similar to human vision in some ways, but we don鈥檛 really know why,鈥 says Murray. 鈥淲e still don鈥檛 have a clear, concise explanation of how either system works.鈥
The researchers say the broader takeaway is clear. These illusions may not be simple mistakes. Instead, they may be a natural result of a system, human or artificial, trying to make sense of the world using incomplete information. What looks like an error may be a reasonable guess based on the visual cues available.
The study also hints at why human vision, and perhaps good models of it, may occupy a middle ground. A system that is too simple fails to capture the complexity of the real world; one that is too powerful, may stop being fooled altogether.
More broadly, the work opens up a new way of studying perception where visual illusions are not just curiosities, but can serve as clues to how both humans and machines make sense of what they see.
Murray hopes AI-based models of perception and cognition could lead to practical applications, from a better understanding of the human mind to new ways of treating psychological disorders.
鈥淚 hope this work is a starting point for explaning human perception much more deeply and thoroughly than we鈥檝e been able to do so far,鈥 says Murray.