Part-to-whole effects and configural processing in faces

How do we recognize a face? One long-running idea holds that faces are processed as wholes rather than as collections of separate pieces. Tanaka and Farah (1993) captured this with what became known as the holistic face effect: a facial feature is easier to recognize when it sits inside the complete face than when it appears on its own. We set out to test how far that effect reaches, and whether it behaves the same way for faces that differ in their individual features versus faces that differ in how those features are arranged.

To do that we ran two experiments with a learning paradigm. The faces came in two varieties. One set varied by componential properties, meaning the individual parts themselves differed (call these COMP faces). The other set varied by configural properties, meaning the spacing and arrangement of the parts differed while the parts stayed similar (CONF faces). Separating the two lets us ask whether holistic processing and configural processing are really the same thing or two distinct contributions to recognition.

In the first experiment, participants learned the full faces. Here the classic pattern came back cleanly: recognition was better when the test showed the whole face rather than an isolated part. So far, nothing surprising. The holistic advantage replicated.

The second experiment changed what people learned. Instead of whole faces, participants studied only the facial parts. Now the holistic advantage shrank, and in some conditions it flipped the other way. That tells us the whole-face superiority is not an automatic property of the stimulus. It depends on whether learning happened holistically in the first place, a point we develop further in Leder and Carbon (in press, 2005).

This fits neatly with the encoding specificity principle of Tulving and Thomson (1973): memory retrieval works best when the conditions at test match the conditions at encoding. Learn the parts, test with parts, and you lose the benefit of the whole. The face data behave like a special case of a much more general memory regularity.

Both the COMP and CONF faces showed these learning-dependent effects, which argues that holistic and configural processing are not interchangeable labels for one mechanism but two separable aspects of how faces are recognized.

One more result is worth laying out. When faces were turned upside down at test, the usual inversion cost appeared for the CONF faces, as expected, since arrangement is exactly what inversion disrupts. But the cost also showed up for the COMP faces, the ones that differed only in their individual parts. That points to some degree of configural processing even for faces defined by their components, blurring the line between the two categories we started with.