Vision, Psychophysics and Bayes

Bayesian decision theory provides a framework within which to develop and test predictive quantitative theories of human visual behavior. Within this framework, we distinguish mechanistic and functional levels in the modeling of human vision. At the mechanistic level, traditional signal detection theory provides tools that can be crucial for drawing correct conclusions about neural mechanisms from psychophysical data through information normalization. At the functional level, the growing development of statistical inference models for neural and perceptual information processing provides for a natural extension of signal detection theory to a pattern inference theory. The key differences between pattern inference and the earlier signal detection theory approach to modeling information is an emphasis on natural tasks and generative models for images and scene structure. We argue that Pattern inference theory is particularly applicable to experimental testing of human visual function. We describe how ideal observers can be used to test theories at both mechanistic and functional levels, and illustrate these two uses with extended examples in motion processing and color constancy.