Conference on Cognitive Computational Neuroscience

Visual cortical populations can be selective for both high-level semantic categories (faces, buildings) and low-level visual features (orientations, spatial positions). The relationship between feature and category selectivity may reflect the distribution of features in natural images, such that neurons are tuned for low-level visual features diagnostic of their preferred category. We evaluate the generality of this hypothesis by asking whether the statistical association between low-level visual features and high-level semantic categories provides a link between feature, spatial, and category selectivity in human visual cortex. Using a large-scale fMRI dataset (Allen et al., 2021) and a voxelwise encoding model based on Gabor features, we find that the distribution of feature and spatial selectivity across voxels within place-, face-, and body-selective ROIs is consistent with the hypothesized roles of these ROIs in high-level visual processing: voxels tend to be tuned for both features and spatial positions that are informative for discriminating their preferred category. These findings suggest that category selectivity in the human brain may reflect the contributions of multiple, hierarchically organized feature spaces.