The role of priors in Bayesian models of perception

In a recent opinion article, Pellicano and Burr (2012) speculate about how a Bayesian architecture might explain many features of autism ranging from stereotypical movement to atypical phenomenological experience. We share the view of other commentators on this paper (Brock, 2012; Friston et al. , 2013; Van Boxtel and Lu, 2013) that applying computational methods to psychiatric disorders is valuable (Montague et al. , 2012). However, we argue that in this instance there are fundamental technical and conceptual problems which must be addressed if such a perspective is to become useful. on the Bayesian observer model (Figure 1), Pellicano and Burr speculate that perceptual abnormalities in autism can be explained by differences in how beliefs about the world are formed, or combined with sensory information, and that sensory processing itself is unaffected (although, confusingly, they also speak of sensory atypicalities in autism). In computational terms, the authors are suggesting that the likelihood function is unaltered in autism and that the posterior is atypical either because of differences in the prior, or because of the way in which prior and likelihood are combined. The latter statement is problematic because within the framework of probability theory, the combination of these two components is fixed as determined by Bayes' theorem: they are multiplied. Put simply, a mathematically consistent Bayesian model cannot accommodate a perceptual abnormality in autism that is due to the way in which belief and sensory information, i. e. , prior and likelihood, are combined. Furthermore, if sensory processing is mathematically represented as a likelihood function (as it typically is within Bayesian models), then changes in the prior cannot lead to changes in sensation, as the authors claim (they can only lead to changes in perception).