Many-Two-One: Diverse Representations Across Visual Pathways Emerge from A Single Objective

Abstract

How the human brain supports diverse behaviours has been debated for decades. The canonical account posits distinct pathways in perceptual processing, such as the "what" and "where/how" visual streams, though their developmental origins and interdependency remain contested. Here, we show that families of deep neural network models develop features that accurately predict hours of human neural and behavioural recordings. By using these models as proxies for the brain and systematically probing developmental objectives, we identify two fundamental computations: object and appearance-free motion recognition, which drive the visual processing hierarchy and explain away alternative accounts. Strikingly, a single objective underlies both: they emerge from optimising for understanding world dynamics, with their organisation highly distributed and continuous across cortex rather than segregating into stream-like modes. Our results suggest that the human brains ability to integrate complex perceptual information across seemingly distinct pathways may originate from the single goal of modelling the world.