Abstract

A morphed face image at the midpoint between a typical face and an atypical face tends to be perceived as more similar to the atypical parent than the typical parent (Tanaka, Kremen, & Giles, 1997). One account of this atypicality bias in face perception is provided by the hypothesis that face representations are characterized as basins of attraction in face space. According to this hypothesis, atypical faces should have larger basins of attraction than typical faces since they are farther from the origin of face space where the density of faces is much lower. This hypothesis is tested on a set of graylevel face images. This paper demonstrates that feed-forward models based on principal component analysis, which have accounted for other face perception phenomena such as "other race" effects, cannot account for the atypicality bias. We next examine some of the assumptions about face space made by the attractor field hypothesis, and demonstrate that typical faces are indeed closer to the origin of face space than atypical faces, and the density of faces is greater near typical faces than atypical faces. Finally, an attractor network model of face representations is implemented, and the basins of attraction are examined by presenting morphed faces to the network. The model replicates the atypicality bias when presented with morphed test patterns, and upholds the prediction that "typical" patterns have smaller basins of attraction than "atypical" patterns stored in an attractor network. We propose that the atypicality bias arises from the competitive interactions required to store typical faces in separate basins of attraction.