Next-Generation Computational Models of Human Perception
-----
Human perception supports flexible decision-making in complex and uncertain environments.
Although modern deep neural networks now achieve human-level performance on many realistic
perceptual tasks, accuracy alone does not constitute an explanation of perceptual
behavior. In this talk, Dr. Govindarajan presents a research program that uses a bidirectional
dialogue between cognitive science and artificial intelligence to develop computational
explanations of human perception. In the first part of the talk, he uses insights
from human behavior to analyze neural network models, showing that they require flexible
computational budgets to learn and deploy generalizable perceptual strategies rather
than relying on task-specific memorization. In the second part, he introduces methods
for representing perceptual uncertainty in deep neural networks and uses these models
to generate quantitative predictions of human perceptual confidence, which are tested
across a range of tasks. Together, these studies illustrate how the confluence of
insights from human behavior and carefully constrained AI models can advance computational
explanations and deepen our scientific understanding of human perception.
-----