Human perception and motor control are often explained as optimal statistical inferences that are informed by accurate prior probabilities.
I will present analyses of human cognitive predictions (Griffiths and Tenenbaum, 2006) and memory retrieval processes (Anderson, 1990) which show that these more central cognitive capacities also approximate optimal statistical inferences given the structure of natural information environments.
The case of memory retrieval offers an elegant demonstration of how a rational statistical framework can be used to span multiple levels of analysis, from computational theory through behavior to neurophysiology.
This is an accomplishment typically only seen in lower-level sensorimotor models.
The Bayesian analysis of people's predictions contrasts interestingly with Bayesian analyses of perception and motor control. While Bayesian analyses in the sensorimotor domain tend to emphasize the discovery of a single prior probability function that characterizes natural environments, Bayesian analyses of cognitive predictions require the use very different forms of priors for qualitatively different classes of events in the world. People's judgments appear to reflect near-optimal use of appropriate priors, in a way that is adaptive to both the qualitative form of the prior and the precise numerical parameters that best characterize empirically observed probability distributions for a particular quantity or event type.