Gender categorization of human faces provides an ideal testing ground for our method, as it has been extensively studied in the past both with psychophysical (Dupuis-Roy et al.,
2009; Mangini & Biederman,
2004; Sekuler, Gaspar, Gold, & Bennett,
2004; Smith, Gosselin, & Schyns,
2004), neuro-imaging (Ng, Ciaramitaro, Anstis, Boynton, & Fine,
2006; Smith, Fries, Gosselin, Goebel, & Schyns,
2009; Smith et al.,
2004) and computational approaches (Abdi, Valentin, Edelman, & O'Toole,
1995; Gray, Lawrence, Golomb, & Sejnowski,
1995; Lu, Plataniotis, & Venetsanopoulos,
2003; O'Toole, Vetter, Troje, & Bulthoff,
1997; O'Toole et al.,
1998). Most computational studies have focussed on building algorithms which can be used to determine the gender of a face from its statistical properties alone, and which thus constitute candidate mechanisms for the algorithms that underlie human gender classification. One influential idea is the observation that the first principal component of (unnormalized) faces is informative about gender (Abdi et al.,
1995; O'Toole, Abdi, Deffenbacher, & Valentin,
1993; O'Toole et al.,
1998; Sirovich & Kirby,
1987; Turk & Pentland,
1991; Valentin, Abdi, Edelman, & O'Toole,
1997). To evaluate the algorithms described above and their plausibility as models of human face processing, their overall performance (in percentage correct) is compared against the performance of human observers (Blackwell et al.,
1997; Hancock, Bruce, & Burton,
1998). We extend these studies by trying to build algorithms which not only predict the over-all percentage correct of human observers, but also their responses on a stimulus by stimulus level (Graf, Wichmann, Bulthoff, & Scholkopf,
2006). Secondly, we work with stimuli for which size, mean pixel intensity and variance of pixel intensity has been normalized, and for which we show the first principal component not to be a good predictor of gender (see
Discussion).