Abstract
Subjects performed a forced-choice, match-to-sample task in which one face image was identical to the sample and the other face differed on a dimension of expression (happy-unhappy), gender, or individuation (two androgynous appearing faces of different individuals each with neutral expressions). The separation of the matching and nonmatching faces was varied by staircase method to yield a 75% threshold. Normal observers were most sensitive to expression changes, followed by gender changes, and least sensitive to identity changes. The threshold stimulus energy for a prosopagnosic, MJH, performing this task was within the range of normal controls when the faces differed along the expression dimension, was at the upper end of the normals when the faces differed in gender, and required much greater stimulus differences when the faces differed in individuation. Adding noise markedly increased the energy requirements of MJH on all tasks, so that he required much greater stimulus differences, relative to the controls, to achieve threshold accuracy, even when the faces differed in expression. According to linear amplifier models, this hypersensitivity to noise for MJH leads to the paradoxical inference that MJH, surprisingly, has very low internal noise compared to controls. Although the data can be fit with the model by assuming MJH is a highly inefficient observer (e.g., he uses a poor template), previous results using reverse correlation techniques suggest this is not the case for his gender and expression discriminations. We explore a model of representation based on the covariance between face images, which captures the population statistics of our sample. Internal noise in the covariance matrix may explain both the sensitivity differences across the three categorizations for normals and MJH, as well as the differences in performance between normal observers and MJH.