The example cell in Figure 8C increased firing rate when aspect ratio dimension was modified but not when the intereye distance changed (Figure S7A). To determine whether cells were significantly tuned for each one of the 19 geometrical feature

dimensions, we repeated the analysis described in Freiwald et al. (2009) and computed GSK3 inhibitor the heterogeneity index (Figure S7B, see Experimental Procedures). Out of the 35 face-selective cells, 29 were modulated by at least one geometrical feature (Figures 8D and S7C), where the most common feature was aspect ratio (Figure S7D). Cells were also modulated by contrast polarity features (Figure 8D). Out of the 35 cells, 19 were modulated by at least one contrast polarity feature. Overall, 49% of the cells were modulated by both types of features (Figures 8E and S7E). Thus, tuning to low-spatial frequency coarse contrast features and to high-spatial frequency geometrical features can co-occur in face-selective cells, suggesting that some cells encode information relevant for both detection and recognition. One of the most FG-4592 solubility dmso basic questions about face-selective cells in IT cortex is how they derive their striking selectivity for faces. Motivated by computational models for object detection

that emphasize the importance of features derived from local contrast (Lienhart and Jochen, 2002, Sinha et al., 2006 and Viola and Jones, 2001), this study focused on the question of whether contrast features are essential for driving face-selective cells. Our main strategy was to probe cells with a parameterized stimulus set, allowing manipulation of local luminance in each face part. The results suggest that detection of contrast features is a critical step used by the brain to generate face-selective responses. Four pieces

of evidence support until this claim. First, different combinations of contrasts could drive cells from no response to responses greater than that to a real face. Second, the polarity preference for individual features was remarkably consistent across the population in three monkeys. Third, the contrast feature preference followed with exquisite precision features that have been found to be predictive of the presence of a face in an image; these features are illumination invariant, agree with human psychophysics (Sinha et al., 2006) and fMRI studies (George et al., 1999 and Gilad et al., 2009), and are ubiquitously used in artificial real-time face detection (Lienhart and Jochen, 2002 and Viola and Jones, 2001). Finally, the tuning to contrast features generalized from our artificial collage of parts to real face images. Shape selectivity in IT has been proposed to arise from cells representing different feature combinations (Brincat and Connor, 2004, Fujita et al., 1992, Tanaka, 2003 and Tsunoda et al., 2001).