, 2009, Walther et al., 2009 and MacEvoy and Epstein, 2011), but they are still far smaller than the likely representational capacity of the human visual system. Theoreticians have argued that the simple statistical properties of natural scenes explain selectivity to low-level features in peripheral sensory areas (Olshausen and Field, 1996 and Smith and Lewicki, 2006). Behavioral data suggest that low-level natural scene statistics also influence the perception of scene categories (Oliva and Torralba, 2001 and Torralba and Oliva, 2003). Though several qualitative theories have

been proposed that link the object statistics of natural scenes with human scene perception (Biederman, 1981 and Palmer, 1975), none have provided an objective, quantitative framework to support this link. The current study provides such a framework. Our data-driven, model-based approach shows that scene categories encoded in the human brain can be derived www.selleckchem.com/products/gdc-0068.html from the co-occurrence statistics of objects in natural scenes.

This further suggests that the brain exploits natural scene statistics at multiple levels of abstraction. If this is true, then natural scene statistics might be used as a principled means Selleckchem Nintedanib to develop quantitative models of representation throughout the visual hierarchy. The work reported here could be extended in several ways. For example, although the spatial distribution of objects within a scene appears to influence the representation of the scene (Biederman et al., 1982, Green and Hummel, 2006 and Kim and Biederman, 2011), the modeling framework used here whatever makes no assumptions about the spatial distribution of objects within scenes. More sophisticated models that incorporate spatial statistics or other mediating factors such as attention may provide further

information about the representation of scenes and scene categories in the human brain. The experimental protocol used was approved by the UC Berkeley Committee for the Protection of Human Subjects. All fMRI data were collected at the UC Berkeley Brain Imaging Center using a 3 Tesla Siemens Tim Trio MR scanner (Siemens, Germany). For subjects S1, S3, and S4, a gradient-echo echo planar imaging sequence, combined with a custom fat saturation RF pulse, was used for functional data collection. Twenty-five axial slices covered occipital, occipitoparietal, and occipitotemporal cortex. Each slice had a 234 × 234 mm2 field of view, 2.60 mm slice thickness, and 0.39 mm slice gap (matrix size = 104 × 104; TR = 2,009.9 ms; TE = 35 ms; flip angle = 74°; voxel size = 2.25 × 2.25 × 2.99 mm3). For subject S2 only, a gradient-echo echo planar imaging sequence, combined with a custom water-specific excitation (fat-shunting) RF pulse was used for functional data collection. In this case, 31 axial slices covered the entire brain, and each slice had a 224 × 224 mm2 field of view, 3.50 mm slice thickness, and 0.