Human Vision Has Quirks That AI Can’t Match

Our eyes do not always represent the world exactly as it is. We are tricked by optical illusions. But some of these imperfections may be adaptive features of biological vision rather than flaws.

Today’s AI vision systems are imperfect in a different way. They are very good at accurately mapping an object in physical space, but they can’t reproduce the ways that picture might be altered by the human visual experience. Read more: “How Your Brain Fills in the Blanks with Experience” Now, a new study from York University researchers suggests that if we want to make AI computational systems more like biological brains, these systems should be able to replicate the perceptual mistakes the human brain makes.

One illusion humans consistently experience is known as the “motion aftereffect.” After we stare at an object moving progressively in one direction, a still object can seem to be slightly offset in the opposite direction. This effect is similar to what happens when you step off of a rocking boat onto dry land and the world continues to wobble temporarily. The perceptual artifact may be a function of efficiency: The brain begins to anticipate movement and must adjust quickly when it cuts out.

To better understand the gap between biological and artificial vision, the team of scientists tested 79 humans, two macaques and nine AI vision networks on this motion aftereffect using a series of moving and stationary images. In their analysis of the results, they found evidence to suggest that primate vision systems may respond to the motion aftereffect in a way that is similar to that of humans, but that AI visual systems do not, even when nudged in the right direction. The effect the scientists found was very small, but Kar and his colleagues suggest some specific ways to build the effect into AI systems through training and feedback.

Neuroscience gives us a way to discover those computations and, potentially, build them into AI.” The training proposals weren’t tested and so are purely hypothetical, and the scientists also did not show that adding in this particular capability to an AI vision network would do anything to improve its performance. But the general idea has plenty of potential applications: More human-like AI visual systems could serve as models for neuroscientific testing, and would better be able to anticipate human perception, which could improve things like medical image displays, driver assistance interfaces or augmented reality. On the other hand, an AI model that can predict human perceptual errors won’t necessarily have human-like values or be easier to control, and so could also be used to design better strategies for deceiving and manipulating humans.

And AI systems are already pretty good at that. Although it’s still unclear whether creating artificially intelligent models that precisely replicate humanity is a path of peril or promise, if we do want AI to be like us, it may help to make it see like us. Subscribe to our free newsletter.