Leonard E. van Dyck^1,2, Katharina Dobs^1,2; ¹Justus Liebig University Giessen, ²Center for Mind, Brain and Behavior, Universities of Marburg, Giessen, and Darmstadt

Category-selective responses are a hallmark of human visual cortex, most prominently for faces, bodies, and scenes. However, it remains unclear how widespread category-selective processing is across object space and which object properties best account for it. To address these questions, we analyzed fMRI responses in inferior temporal cortex (ITC) of three participants viewing images from hundreds of object categories in the THINGS database. Using a shared response model, we derived a common feature space across participants and introduced a sparseness-based selectivity metric quantifying how strongly each category engages selective features. This approach confirmed robust selectivity for faces, bodies, and scenes under broad category sampling, but also revealed previously unknown selectivities. Applying the same framework to high-level representations of deep neural networks (DNNs), we found only moderate correspondence between category selectivity rankings in ITC and in DNNs. To identify the object properties underlying these rankings, we fit an encoding model using behavior-derived object dimensions to predict each category’s selectivity. In ITC, category selectivity was explained by a small set of primarily semantic dimensions, whereas in DNNs it was captured by a larger set of mostly visual dimensions. Together, our findings show that category-selective processing extends across object space beyond classical domains and is organized by interpretable, feature-based principles. These principles point to a fundamental dissociation between biological and artificial visual systems: visual cortex prioritizes ecologically important categories, whereas current computational models emphasize visually distinctive ones.