A reverse prediction test shows that supposedly brainlike AI models may depend on visual strategies the primate brain does not use.
Artificial intelligence can sometimes predict how the brain responds when people recognize objects. But that resemblance may hide an essential weakness: the internal workings of today’s vision models do not necessarily fit the processes utilized by a primate brain.
Over the past decade, synthetic neural networks (ANNs), computer models designed for visual tasks, have become to be some of the main tools for explaining how the brain processes sight. York University researchers wanted to determine whether or not these systems genuinely function like biological vision.
“Artificial intelligence systems are often explained as ‘brain-like’ because they can predict activity in parts of the brain that support us recognize objects,” stated York University Assistant Professor Kohitij Kar, senior author of a new study. “Until now, scientists mostly examined this in a one direction. They asked whether AI models can predict brain activiy.”
The researchers reversed that acquainted test. If AI genuinely displays the brain, they reasoned, then recorded mind activity ought to also expect the model’s inner responses. To study this possibility, they developed a opposite predictivity test.
“Ultimately, we require computational models to clearly understand the underlying neural mechanisms of how we understand objects. How can we see objects move? While it’s a very easy task that we do everyday, computationally, although, it’s a very challenging issue,” says Kar, the Canada Research Chair in Visual Neuroscience and a member of York’s Centre for Vision Research and Centre for Integrative and Applied Neuroscience.
A reverse test challenges brainlike AI
The researchers, including York Postdoctoral Fellow Sabine Muzellec, a Connected Minds trainee, examined the models with 1,320 herbal photographs and realistic synthetic snap shots. The set included bears, elephants, faces, apples, motors, puppies, chairs, planes, birds, and zebras shown against indoor, out of doors, and other natural backgrounds.
They also used 300 additional images depicting the same objects in altered forms, consisting of outlines, drawings, simplified representations, and artist variations. This wider range helped test whether or not the relationship among brain activities and AI features held across different visual styles.
Brain activity exposes a hidden mismatch
“The outcomes have been striking. While AI models can expect the neurons we recorded in the brain fairly well, the brain can not equally expect the various model’s internal features. Interestingly, this is not the case when neurons from one brain is compared against ones from another brain,” stated Kar.
This imbalance shows that ANNs may attain correct visual answers via processes that differ from those used by primate brains. Kar warns that the mismatch could broaden as models become more complex until researchers address it early. If an AI model anticipates neurons however its own internal features cannot be predicted from neural activity, it may no longer offer a dependable explanation of how the brain works.
“The findings suggest that these days’s AI systems solve visual tasks in partly internal strategies that the brain may not use. Importantly, the parts of AI models that align with the brain are also better at predicting real human behavior,” says Kar.
Why this matters
Researchers increasingly use AI models to design studies of human behavior, together with clinical research. Much of that work assumes the systems process the world in ways that resemble the human mind.
“Our findings challenge how similar current AI systems definitely are with the primate brain. We show that models that have been previously thought to be brain-like depend upon internal systems that the brain does not appear to use. We give a well-vetted diagnostic metric for the field,” says Muzellec.
Better alignment ought to reinforce research
More correct brain-like models could eventually support research related to conditions ranging from post-traumatic stress disorder to autism. For now, moreover, using poorly aligned systems to interpret human behavior carries risks. Comparable models are also being applied to hearing, language and movement, making dependable validation important across numerous fields.
“Our technique supports identify which parts of an ANN in reality match brain activity, permitting us to build more reliable models for understanding how people see and interpret the world,” says Kar. “This is specifically vital for our autism research program, which builds on models of the neurotypical brain as a baseline.”
