Bayesian approaches to perception
Randy L. Diehl, Sarah C. Sullivan · The Journal of the Acoustical Society of America · 2003
The concepts of Bayesian statistical decision theory have recently transformed research in perception by providing a rigorous mathematical framework for representing the statistical properties of the environment, for describing the tasks that perceivers perform, and for deriving computational theories of optimal performance on those tasks. Such computational theories (ideal observers) offer an appropriate benchmark for evaluating human performance, and, moreover, they can be modified to yield an excellent starting point for developing testable models. Unlike theories that focus on the role of invariant information for perceived events, the Bayesian approach treats information as probabilistic. The approach is illustrated for cases in which listeners learn to categorize artificial stimulus sets whose statistical properties (prior probabilities and stimulus likelihoods) are controlled by the experimenter. Prospects for extending the Bayesian framework to the analysis of speech perception are discussed in light of the recent progress in developing ideal observers for natural environmental stimuli. A necessary but daunting task will be the measurement of probability distributions that characterize natural speech categories. [Work supported by NIDCD.]