Ideal Observers, Real Observers, and the Return of Elvis
Ronald A. Rensink, Vision Sciences Laboratory, Department of Psychology, Harvard University, Cambridge, Mass. USA.

In DC Kersten and W Richards (eds.), Perception as Bayesian Inference (pp. 451-455). Cambridge: Cambridge University Press. 1996.   [pdf]
[Commentary on Knill DC, Kersten D, and Mamassian P (1996) "Implications of a Bayesian Formulation of Visual Information for Processing for Psychophysics". In Perception as Bayesian Inference, DC Knill and W Richards, eds.]

Abstract

Knill, Kersten, & Mamassian (Chapter 6) provide an interesting discussion of how the Bayesian formulation can be used to help investigate human vision. In their view, computational theories can be based on an ideal observer that uses Bayesian inference to make optimal use of available information. Four factors are important here: the image information used, the output structures estimated, the priors assumed (i.e., knowledge about the structure of the world), and the likelihood function used (i.e., knowledge about the projection of the world onto the sensors). Knill et al argue that such a framework not only helps analyze a perceptual task, but can also help investigators to define it. Two examples are provided (the interpretation of surface contour and the perception of moving shadows) to show how this approach can be used in practice.

As the authors admit, most (if not all) perceptual processes are ill-suited to a "strong" Bayesian approach based on a single consistent model of the world. Instead, they argue for a "weak" variant that assumes Bayesian inference to be carried out in modules of more limited scope. But how weak is "weak"? Are such approaches suitable for only a few relatively low-level tasks, or can they be applied more generally? Could a weak Bayesian approach, for example, explain how we would recognize the return of Elvis Presley?

It is argued here that a weak Bayesian approach is suitable only for a task that avoids ill-defined structures and resource-limited processes, and has well-defined priors that are relatively invariant, at least under some sets of conditions. But many perceptual problems (such as recognizing the return of Elvis Presley) are not of this type, and Bayesian analyses of such tasks therefore cannot succeed.

Back to main publications list.