Abstract
According to the face-space framework, face representations are isomorphic to locations in a multidimensional psychological space, in which the distance separating these representations is proportional to the degree of dissimilarity between faces. This similarity-based face-space has empirically accounted for a range of face-processing phenomena. Our aim was to test whether the similarity-based structure of face-space could also mediate identity invariance – the fundamental ability to maintain constant identity representations under varying face transformations (e.g., changes in lighting or viewpoint). This invariance can be achieved if similarity relations remain unchanged across different transformations. We therefore examined the extent to which similarity relations among faces were indeed constant under view or lighting transformations. In Experiment 1, subjects rated perceived similarity within a set of facial stimuli, viewing either its frontally-lit variant or its top-lit variant. Two group-averaged face-space configurations were constructed from these ratings, and their degree of concordance was estimated using Procrustean analysis. In Experiment 2, subjects rated perceived similarity both for a frontal-view variant and a 60°-view variant of the same stimuli set, in two separate sessions three weeks apart. Concordance was estimated both for inter-subject and intra-subject spaces. Consistent with our hypothesis, the fit between spaces constructed for different views or lighting transformations was significantly high, indicating that similarity relations were kept constant under these transformations. Furthermore, multidimensional spaces created for relatively similar transformations (e.g., frontal-lighting space and frontal-view space) showed higher concordance than those created for more distant transformations (e.g., top-lighting and 60°-view). Finally, intra-subject spaces were found to be more in accordance with each other than inter-subject spaces, suggesting that similarity across group-averaged spaces was not due to averaging. Overall, our findings suggest that invariant identity processing can be achieved by keeping the distance between face exemplars in face-space similar under different transformations.