Movellan, Javier
UCSD
Institute for Neural Computation

The Morton-Massaro law: Analysis and implications for models of perception

Many experiments and demonstrations document the fact that the perception of a wide range of stimuli is affected by contextual information. Guided by a model that he called the logogen, John Morton predicted that in experiments that require identifying one of two alternatives the ratio of the choice probabilities of these alternatives should factorize into components independently controlled by stimulus and context. These components are interpreted as the relative support of the stimulus and context for the two alternatives under consideration. Massaro and colleagues have shown that this form of factorability of information sources provides good approximations to empirical response probabilities obtained in a remarkable range of domains such as word and letter perception, object identification, depth perception, memory retrieval and recognition of emotions. It has also proposed that this pattern of results is incompatible with feed back models of perception. We analyze under what conditions the pattern of results observed by Morton and Massaro reflects optimal perceptual inference. We also show that factorability is not incompatible with feed-back models and it is simply an indication of an architectural constraint that we named "channel separability". Experiments assessing factorability of different sources of information may be useful in diagnosing the structure of a processing system. We illustrate this with an experiment demonstrating a violation of factorability predicted by a feed-back model of perception.