Large-Scale In Silico Analysis of Visual Property Representations Across Cortical Regions and Representational Metrics

Visual neuroscience often faces a tradeoff between experimental control and stimulus richness: tightly controlled studies typically probe a small number of properties, whereas natural-image datasets provide scale and realism, but contain many entangled sources of variation. We introduce a large-scale in silico framework that combines generative models with image-to-fMRI prediction to systematically study controlled visual property variation in naturalistic face and body images. The generated stimulus space spans more than 80 properties across hundreds of identities, enabling matched comparison of the same controlled manipulations across multiple visual cortical regions and four subject-specific predicted response profiles. We find broad but heterogeneous property representation that recovers expected face- and body-related cortical organization. We further compare five complementary analysis families (sensitivity, cross-identity representational geometry, linear decoding, coding-direction consistency, and property-wise encoding), and find they agree only partially. In particular, we find that linear decodability appears to track response sensitivity much more strongly than identity-general geometry, while stable structural properties often exhibit conserved geometry without commensurate levels of linear decodability. Together, this framework provides a simultaneously scalable and granular handle on visual representation mapping and suggests that conclusions about what is "represented" depend strongly on the representational question we are asking, as well as how we are asking it.