Face Perception

Face Recognition

Humpty Dumpty's complaint illustrates one of the central challenges in recognizing individual faces. Unlike many other visual objects, we do not distinguish faces by their parts alone or even by the basic arrangement of these parts. Rather, successful face recognition depends on sensitivity to subtle spatial relations between facial features, sometimes referred to as configural coding. Unlike most other objects (except perhaps words), faces are also coded holistically, with little explicit coding of component features. Turning faces upside down disrupts recognition of faces far more than other mono-oriented objects, and this disproportionately large effect is likely due to inversion dramatically reducing our sensitivity to spatial relations (see Figure 1) and disrupting our holistic coding of faces.

Figure 1 Face recognition test

Some faces are easier to remember than others, and this may be a consequence of how the similarities and differences among faces are represented. Individual faces can be thought of as “points” in a multidimensional space. At the center of face space lies an average face, representing the central tendency of all the faces experienced. Typical faces, which are harder to remember, are located close to this average, while more distinctive faces, which tend to stick in our memory, are located further away (see Figure 2). Recent evidence suggests that the average face functions as a perceptual norm for coding [Page 344]identity, highlighting what is distinctive about each face. The norm appears to be continuously updated by experience. A curious but telling side effect of this updating is that viewing a face biases us to see the “opposite” face, the so-called identity aftereffect (see Figure 2). Norm-based coding may be implemented by pairs of neural populations, one tuned to above-average and one to below-average values, for each dimension of face space. Common or typical faces produce little neural response with large responses reserved for novel faces, resulting in an efficient form of coding that may be used more generally in the visual system.

...