Why are “What” and “Where” Processed by Separate Cortical Visual Systems? A Computational Investigation

TLDR

Primate vision separates object identification from spatial processing across distinct cortical pathways. The study investigates the computational properties of a two‑system design by building simplified connectionist models. The authors built networks that simultaneously classify and locate shapes in a matrix, comparing fully connected architectures to ones with hidden units split into identity and location groups. Simulations showed that separate identity and location streams improved performance only when sufficient resources were available, enabling more efficient internal representations.

Abstract

In the primate visual system, the identification of objects and the processing of spatial information are accomplished by different cortical pathways. The computational properties of this "two-systems" design were explored by constructing simplifying connectionist models. The models were designed to simultaneously classify and locate shapes that could appear in multiple positions in a matrix, and the ease of forming representations of the two kinds of information was measured. Some networks were designed so that all hidden nodes projected to all output nodes, whereas others had the hidden nodes split into two groups, with some projecting to the output nodes that registered shape identity and the remainder projecting to the output nodes that registered location. The simulations revealed that splitting processing into separate streams for identifying and locating a shape led to better performance only under some circumstances. Provided that enough computational resources were available in both streams, split networks were able to develop more efficient internal representations, as revealed by detailed analyses of the patterns of weights between connections.