Maithilee Kunda, 4:00pm Monday June 18th
Vanderbilt University
Monday, June 18th
Swift 107
4:00pm (Reception to follow)
Despite evidence for the importance of visual mental imagery in many areas of human intelligence, including reading, mathematics, creativity, scientific discovery, computer programming, and more, there are few AI systems that use imagery-like knowledge representations to perform complex tasks. Part of the problem comes from confusion in the AI literature between (1) tasks that are presented visually, in an external visual format, versus (2) tasks that are solved visually, using internal visual representations.
There is a rich history of AI research in the first category, but the vast majority of these AI systems first convert visual inputs into internal propositional (i.e. abstract/symbolic) representations before solving a task. Fewer studies fall into the second category, but these studies do provide insight into the computational nature of imagery-based representations and their potential roles in intelligence. I present results from our research that include: how imagery-based representations and reasoning operators can be combined to solve standardized tests of nonverbal cognition, like the Raven's Progressive Matrices test; how imagery-based reasoning operators, like mental rotation, can be represented in connectionist systems and learned from perceptual experience; and how educational technologies that leverage visual cognition can be used to help people with autism or other atypical cognitive conditions.