Abstract
Many authors distinguish "first-order" object recognition from "second-order" tasks that are poorly suited to template matching and seem to demand other kinds of perceptual computation. "Second-order" tasks include detecting symmetry, Glass patterns, modulations of white noise, and coarse patterns composed of small balanced elements. Past treatments have suggested various special computations, particular to each task, that observers might make. We take a more general approach. We suppose a complete set of receptive fields (like those of V1 cells) and ask how many receptive fields are required to perform as well as human observers. This is like defining efficiency as the fraction of available information (e.g. dots or area) that would be required by an ideal observer, but applied to receptive fields rather than to components (e.g. dots) of the stimulus. With mild assumptions about the receptive fields, this reveals a dichotomy between "first-order" ordinary identification tasks that require on the order of ten receptive fields and "second-order" tasks that require thousands or millions. The necessary cortical wiring is greatly affected by the hundred-or-more-fold increase in the number of receptive fields used.
Meeting abstract presented at VSS 2012