Abstract
It is widely held that the neural representation of faces may be usefully viewed as a multi-dimensional “face-space” with the average at its origin. Under this view the distance between faces encodes their similarity, and distance from the origin encodes general “distinctiveness”. Here we compare objective and subjective measures of distinctiveness to uncover the nature of the dimensions supporting it. To this end we measured key-points for 64 Caucasian (Scotch) male faces, to yield a 70-parameter “identity vector” (IV) for each. Our stimuli were average-faces morphed to shift them along IVs.
We estimated subjective distinctiviness by having observers rate 16-member subsets of the face-set. We then made two objective measures of distinctivness: IV-length (i.e. distance from average) and psychophysical discriminability of the face from the average. The latter was the smallest shift of an average face along the IV (estimated using QUEST) that supported a 3AFC odd-man-out judgment (i.e. the cued versus two averages). We report that subjective ratings correlate well ([[gt]]85%) with discriminability of the face from average, but less so with IV-length, suggesting that such thresholds are a useful psychophysical measure of distinctiveness. We next measured discrimination thresholds for faces morphed along the IV that were presented not just amongst averages, but amongst averages containing varying “pedestal” levels of morphing along the IV. Results indicate first, broadly inverse Weber's law dependence of threshold on pedestal; subjects get better at spotting the odd-man-out as faces become more atypical. Second, some faces elicit shallow “dipper functions” indicating discrimination is based on a non-linear transduction of (IV-encoded) identity level. By contrast, performance with inverted faces was uniformly poorer and exhibits straightforward inverse Weber's law behaviour with no dips. Thus our findings are related to encoding of facial distinctiveness and not visibility of distortion.
Funded by the Wellcome Trust