Representational Geometries Across Visual Working Memory Encoding and Maintenance

Abstract

This study examined the representational geometry of visual working memory using human 7T fMRI. Observers viewed a facescene blended sample, attended to either the face or scene, and judged whether the attended aspect matched a subsequent test image. We found that the coding dimension distinguishing faces and scenes rotated between encoding and maintenance phases, in both visual and association areas. In these regions, exclusive combinations (i.e., encoded face and maintained scene vs. encoded scene and maintained face) were linearly decodable, indicating encoding and maintenance phases represent visual information using distinct subspaces. Such high-dimensional, flexible geometry can, in principle, protect maintained visual information from incoming input. At the same time, robust crossdecoding was observed across phases, reflecting the stability and generalizability of the represented contents. The balance between representational flexibility and stability varied along the cortical hierarchy: early sensory regions emphasized flexibility, whereas transmodal regions showed greater stability.