A Dedicated System for Processing Faces

If you're planning to rob a bank, there's one thing you must not forget: to cover your face. Otherwise, just a brief glance will allow all the other social animals around you to identify you. What is the neural basis of the extraordinary ability of humans to recognize faces? Localized strokes can selectively destroy face recognition abilities while preserving the ability to recognize other objects ("prosopagnosia") (1). Furthermore, functional magnetic resonance imaging (fMRI), a technique that measures blood flow changes induced by brain activation, consistently reveals several discrete brain regions that respond more to faces than to other objects (2). One of these regions, the fusiform face area, shows increased blood flow even when subjects merely imagine faces (3). These findings suggest that face processing is mediated by specialized modules inside the human brain. Such specialization is surprising since, from introspection, it seems that our recognition of faces flows seamlessly into that of all the other objects in the world.

Are face-selective regions unique to humans? Charles Gross and co-workers studied a large region in the macaque brain known as the temporal lobe and reported in 1981 that this region contains some cells that respond exclusively to faces and not to other visual forms (4). This was a remarkable finding: How can a single cell be wired to detect something so complex as a face? The discovery immediately turned fuzzy questions about holistic integration and gnostic units into a concrete research goal: What are face cells detecting, and how are they wired?

One problem, however, stood in the way of understanding these cells: It was difficult to find them. In single-unit recording experiments, one can see only as far as the tip of one's electrode (100 Ð'µm wide). Several groups that studied face cells reported that they were scattered throughout the temporal lobe, with at most 10 to 20% of the cells in any one region being face-selective (5-7). Meanwhile, the discovery by fMRI of face-selective regions in humans generated great interest in understanding what is being coded by these regions. One might guess that the fMRI-identified face-selective regions contain lots of face cells. Alternatively, they could contain cells activated by any animate object, or by symmetrical objects, or by the behavioral process of fine scrutiny. Indeed, fMRI evidence was marshaled for several competing theories about face-selective activation.

In order to clarify the link between face cells and fMRI face areas, I performed fMRI experiments in alert monkeys to look for face-selective regions (8). Comparing activation to faces versus five other object categories (fruits, bodies, gadgets, hands, and scrambled patterns) across the entire macaque brain, I identified face-selective activation in three discrete regions of the temporal lobe (see the figure, left, panel A). These regions showed a blood oxygen level-dependent (BOLD) response to faces that was stronger than the response to any of the nonface categories by a factor of 7 (see figure, panel B). This suggests that face processing in monkeys is performed by specialized regions, possibly homologous to those found in humans. Furthermore, the arrangement of the face regions along an anterior-posterior axis suggests a hierarchy, given that we know from other studies that complexity in shape selectivity increases from the back to the front of the visual system.

Recognizing faces: (A) Three patches of face-selective fMRI activation (yellow regions) in the macaque temporal lobe. (B) Time course from the face patches. Blood flows to these regions only when the monkey views faces. (C) Average response across 182 cells from the middle face patch of one monkey to 96 different images. The first 16 images are faces. (D) Responses of a face cell to repeated presentations of an upright and an inverted cartoon face. Each dot represents an action potential.

Having found these fMRI face regions in monkeys, I then asked: What is the selectivity of single neurons within an fMRI-identified face patch? I started by recording from single neurons--almost 500 of them--in the middle face patch of two monkeys, and found 97% of the visually responsive neurons to be face-selective (9) (see figure, panel C). These cells responded almost 20 times as strongly to faces as to other objects, and many were even suppressed by nonface objects. Up to now, one major difficulty with understanding object recognition