Representations converge as brain maps diverge along the cortical hierarchy

Abstract

Brain maps (e.g. topographies like retinotopy, somatotopy) vary across individuals. This is thought to reflect computational differences, motivating personalized functional localization. However, artificial neural networks (ANNs) show that similar performance and internal representations can coexist with diverse circuit layouts. Consequently, we tested whether spatial diversity reflects computational diversity in the brain. Using task and resting-state fMRI data, we compared regional functional topographies and representational geometries-the within-individual dissimilarities among activity patterns that capture computable information. Across individuals (n = 414), representations converged in higher-order cortex despite substantial topographic diversity. Thus, different, individual-specific activity patterns encoded similar information. Topographic differences only tracked representational differences in regions under strong architectural constraints, such as highly myelinated sensory-motor cortices. We show this parallels ANNs which begin with random initial layouts but learn convergent representations, if architectures are permissive. This parallel raises a developmental question: do topographies and representations also have different developmental origins? Examining twins (n = 394), we found representations were sensitive to developmental environments and less heritable than topographies. Together, this shows that representational convergence occurs across idiosyncratic layouts in both artificial and biological systems, but is moderated by architectural constraints on circuit implementations. Accordingly, the relevance of localization-and representation-based paradigms of brain function depends on neural architecture.