Manipulating Face Gender

Abstract

Previous research has shown that faces coded as pixel-based images may be constructed from an appropriately weighted combination of statistical "features" (eigenvectors) which are useful for discriminating members of a learned set of images. We have shown previously that two of the most heavily weighted features are important in predicting face gender. Using a simple computational model, we adjusted weightings of these features in more masculine and more feminine directions for both male and female adult Caucasian faces. In Experiment 1, cross-gender face image alterations (e.g., feminizing male faces) reduced both gender classification speed and accuracy for young adult Caucasian observers, whereas same-gender alterations (e.g., masculinizing male faces) had no effect as compared to unaltered controls. Effects on femininity-masculinity ratings mirrored those obtained on gender classification speed and accuracy. We controlled statistically for possible effects of image distortion incurred by our gender manipulations. In Experiment 2 we replicated the same pattern of accuracy data. Combined, these data indicate the psychological relevance of the features derived from the computational model. Despite having different effects on the ease of gender classification, neither sort of gender alteration negatively impacted face recognition (Experiment 3), yielding evidence for a model of face recognition wherein gender and familiarity processing proceed in parallel.