Abstract
The inverse relationship between the rated typicality of a face and its recognizability is thought to reflect the properties of the similarity space for human representations of faces. Because most typicality studies use identical images to assess typicality and to test recognition, it is unclear whether the typicality-recognizability relationship reflects similarity in an image representation space or a face representation space. We examined the stability of typicality ratings under different illumination conditions and assessed the typicality-recognizability relationship over changes in these conditions. In Experiment 1, participants rated the typicality of two images of each of 108 faces, with one image taken under controlled, and the other taken under uncontrolled illumination. In Experiment 2, a separate set of participants performed a recognition test, in which the learn-test transfer direction between controlled and uncontrolled illumination images was varied. Performance was better when participants learned controlled illumination images and were tested with uncontrolled illumination images than for the opposite direction of transfer. We found a moderately strong correlation between typicality ratings for different images of the same person. The typicality-recognizability relationship held across varying illumination conditions, but was weaker than that found in previous studies using the same image. It was also weaker than the relationship we found in a control study when illumination conditions remained constant but different images of a face were used. Notably, in all cases, the typicality-recognizability relationship was due primarily to the high false alarm rates associated with more typical faces. These results suggest that part, but not all, of the relationship between face typicality and recognition accuracy reflects properties of the image space rather than the ‘face space’.
Supported by the Technical Support Work Group (TSWG).