Bonn Melbourne Seminar in Decision Making and Computational Psychiatry
Abstract
Classic studies in active vision have long put forward the idea that perception, cognition, and action are difficult
to separate in goal-directed sensorimotor behavior. This has led to an emphasis of the task as explaining visual
behaviors relative to low-level image features. More recently, computational models of sequential decisionmaking
under uncertainty, so-called Partially Observable Markov Decision Processes, have been used to model
such behaviors. These types of models can be seen as a generalization of decision making under uncertainty to
sequential decisions or as a generalization of reinforcement learning type tasks also to include state
uncertainty. In this talk, I will present a series of studies that use probabilistic sequential decision-making
models to understand human visual behaviors. Experiments include monitoring, visual search, and event
detection tasks, as well as sensorimotor tasks such as navigation and ball catching. The recurrent theme will be
that under the full uncertainties, ambiguities, and partial information in naturalistic tasks, perception,
cognition, and action are inseparably intertwined, requiring trial-by-trial and moment-by-moment
computational models to understand human behavior.