A Biased View of PerceiversCommentary on ` Observer theory , Bayes theory

Commentary on`Observer theory, Bayes theory, and psychophysics,' by B. Bennett, et al. 1 Observers and Theorists The target chapter 1 introduces the notion of an observer as an entity which applies a given Bayesian theory to the task of inferring world properties from sense data. In fact, most of this volume is concerned with observers in one form or another. The basic idea is that the perceptual system uses a model of the current world context, that is, a `domain theory,' to reduce the ambiguity inherent in the sense data and to select particular interpretations for the state of the world. Example domain theories are: the qualitative probabilistic models used for the motion of a ball in a box in Chapter 4; the preference based scheme for the interpretation of handle shape and orientation discussed in Chapter 6; the Bayesian models for shape from texture and shading used in Chapter 10; and the cost function formulation proposed for the interpretation of shape, illumination, and pigmentation in Chapter 11. In order to include such alternative forms of world models we use the termòbserver' to refer loosely to any entity which applies a xed domain theory to the task of inferring world properties from sensory data. Our bias is to view the perceptual system as a theorist in its own right, not just an observer. We claim that in order to understand perception we need to understand the way this theorist works, that is: how it selects a theory for a new domain, how it applies a theory to choose particular interpretations for sense data, how it compares diierent theories, and how it revises and learns theories under the innuence of new information. In contrast an observer is concerned with only one of these aspects, namely how a xed theory is applied. The distinction between theorist and observer is perhaps blurred by notions such as Yuille's competitive priors (see Chapter 8 or 8]). Each prior may lead to a single distinct (Bayesian) observer, and the selection of one prior in favor of others can be thought of as a theory selection process. Thus the overall system can be thought of as a theorist. On 1 This commentary is to appear in the book \Perception as Bayesian Inference" edited by D. Knill and W.