Binxu Wang (right) and her former PhD advisor and current collaborator Carlos Ponce used generative AI to direct neurons to characteristics of test images, revealing unexpected patterns in their preferences.
Photo credit: Chris Brewer
How can scientists figure out what visual information neurons actually respond to?For decades, neuroscientists have tried to answer this question by showing pictures of animals, such as faces, trees, houses, and other animals. How neurons in brain areas associated with vision respond to these pictures gives researchers clues about the type of information the neurons are processing from the pictures.
Although this approach has yielded important discoveries, it also has notable limitations. “When you see neurons that respond to cats, it’s tempting to call them ‘cat neurons’ or ‘face neurons,'” says Binshu Wang, a researcher at the Kempner Institute for Natural and Artificial Intelligence. “This can be useful, but it can also be a trap.”
The problem, Wang said, is that the test images were chosen by human researchers. Because researchers rely on their own visual intuition to select and interpret images, they may miss the very patterns to which neurons in the primate visual system respond most strongly. In other words, the images humans select may not fully reflect the kind of information to which neurons preferentially respond.
In a new paper just published in Nature Neuroscience titled “Neuronal tuning dynamically adjusts with object and texture manifolds across the visual hierarchy,” Wang and her former PhD advisor and current collaborator Carlos Ponce, assistant professor of neurobiology at Harvard Medical School and affiliated faculty member in the Kempner Institute, present a new approach to this problem. They use generative AI to force neurons to shape the properties of test images themselves, revealing hidden structure in neurons’ visual preferences.
Source link
