Norm-based face encoding by single neurons in the monkey inferotemporal cortex David A. Leopold 1 †, Igor V. Bondar 1 † & Martin A. Giese 2 The rich and immediate perception of a familiar face, including its identity, expression and even intent, is one of the most impressive shared faculties of human and non-human primate brains. Many visually responsive neurons in the inferotemporal cortex of macaque monkeys respond selectively to faces 1–4 , sometimes to only one or a few individuals 5–7 , while showing little sensitivity to scale and other details of the retinal image 8,9 . Here we show that face-responsive neurons in the macaque monkey anterior infero- temporal cortex are tuned to a fundamental dimension of face perception. Using a norm-based caricaturization framework pre- viously developed for human psychophysics 10–12 , we varied the identity information present in photo-realistic human faces 13 , and found that neurons of the anterior inferotemporal cortex were most often tuned around the average, identity-ambiguous face. These observations are consistent with face-selective responses in this area being shaped by a figural comparison, reflecting struc- tural differences between an incoming face and an internal reference or norm. As such, these findings link the tuning of neurons in the inferotemporal cortex to psychological models of face identity perception. Previous work showing that face recognition benefits from carica- turization—that is, the exaggeration of distinguishing features 10,14–16 — has led to the conceptualization of a multidimensional ‘face space’ (see cartoon in Fig. 1a), within which each point represents an individual face and the average of all faces resides at the centre 17 . We focused on stimuli along four face-space trajectories branching away from the average face (sometimes termed the ‘norm’ or ‘prototype’ face) towards increasing identity levels. This radial axis, sometimes termed an axis of caricaturization or identity trajectory, has been studied extensively, and is thought to play an important role in face percep- tion 10–12,18,19 . Also shown are faces separated tangentially, morphed directly between faces of different identities. The construction of photo-realistic faces along radial and tangential trajectories in a face space were generated with a ‘morphable face model’, which inter- polates between three-dimensional laser scans of human heads 13 .A recent study using these stimuli found that equal increments along the caricaturization axes in a face space defined by such a morphable model translated directly to equal perceptual increments 12 . These data are shown for humans in Fig. 1b, and are compared directly to similar data from one of the monkeys used in this study (N97). This monkey recognized human faces in a manner similar to human subjects (albeit with a slightly diminished performance), and experi- enced a face-identity after-effect in the same way as human subjects 20 (see also Supplementary Fig. 2). To test whether the psychological representation of the caricatur- ization axes was reflected in the firing of single neurons in the inferotemporal cortex, we performed extracellular microelectrode recordings in two monkeys (see Methods) that were presented with a large number of faces corresponding to different positions in face space. The faces were presented in a pseudorandom sequence. The microelectrodes were situated in the anterior inferotemporal cortex (AIT) region (see Supplementary Fig. 3), an area known to contain neurons that respond selectively to face and other complex patterns. Each day, the monkey was presented with faces from four different axes of caricaturization, corresponding to four different individuals. Different sets of faces were used for the two monkeys. One of the animals, C99, was additionally tested with faces modulated along tangential directions of face space (see dotted line in Fig. 1a), corresponding to morphs between different face identities. A total of 250 face-responsive units were monitored, with 209 tested with a complete set of faces for our evaluation. Despite several differences in the training and testing, and in the electrophysiological procedures (see Methods), the results were similar in both animals. The main finding was a striking tendency for neurons to show tuning that appeared centred about the average face. For a large proportion of neurons, this entailed monotonic trends in firing for LETTERS Figure 1 | Face space and behavioural data. a, Each identity trajectory (F1–F4) intersects in the centre, at the average face (0% identity). Identity levels change from the average through to the full identity (100%) faces, to the caricatures (160% shown, asterisks). Other face identities in this study (for example, between F2 and F3, dotted blue line), lay on the tangential trajectories. b, Behavioural performance of monkey N97 in the recognition task, shown for two different face presentation durations and compared to seven humans (mean ^ s.e.m.) performing the same task (with 500-ms presentations). Chance performance for correctly recognizing a face out of four learned identities was 0.25 (horizontal dashed line). Vertical dashed line corresponds to zero identity level. 1 Max Planck Institut fu ¨r biologische Kybernetik, Spemannstrasse 38, 72076 Tu ¨bingen, Germany. 2 Laboratory for Action Representation and Learning, Department of Cognitive Neurology, Hertie Institute for Clinical Brain Research, University Clinic Tu ¨bingen, Schaffhausenstr. 113, D-72072 Tu ¨bingen, Germany. †Present addresses: Unit on Cognitive Neurophysiology and Imaging, Laboratory of Neuropsychology, National Institute of Mental Health, National Institutes of Health, 49 Convent Drive, Bldg. 49, MSC 4400, Bethesda, Maryland 20892, USA (D.A.L.); Institute of Higher Nervous Activity and Neurophysiology, Russian Academy of Sciences, Butlerova Street 5a, 117485 Moscow, Russia (I.V.B.). Vol 442|3 August 2006|doi:10.1038/nature04951 572 © 2006 Nature Publishing Group