Visualizing the global geometry of population representations of multiple visual object categories with spheres

Brain computation can be understood as the transformation of representations across stages of processing. The content and format of a representation is reflected in the geometry of stimulus-related response patterns. Here we characterize the representations along the human ventral visual pathway with a new visualization technique called “hypersphere2sphere” (H2S). It takes as input a labeled set of points in a high-dimensional space (the multivariate response space of each cortical region) and fits a hypersphere to represent each category. It visualizes these high-dimensional hyperspheres as a set of spheres in 3D (or circles in 2D), revealing their relative sizes, separations, and overlaps. Using functional magnetic resonance imaging (fMRI), we measured response patterns to 48 images from four categories (faces, bodies, inanimate objects, and scenes). We computed unbiased distance estimates in representational space using crossvalidation. With H2S, we observed the emergence of response pattern clustering, based on category, at the level of the lateral occipital complex. Categories also occupy nonoverlapping hyperspheres in faceand place-selective areas, with faces most spread in the former and scenes in the latter. H2S provides a useful perspective on highdimensional representational geometries that promises new insights on the basis of hemodynamic and electrophysiological brain-activity measurements.