Working memory operations emerge from dynamic changes in neural subspace geometry

Abstract

Working memory representations of multiple items are maintained within quasi-orthogonal neural subspaces, a mechanism thought to reduce interference between memory contents. Previous studies in both non-human and human primates have demonstrated the existence of such subspaces, but it remains unclear how cognitive operations are performed on the information they encode. One hypothesis proposes that these operations are implemented through dynamic changes in subspace geometry, yet direct empirical evidence for this idea in the context of working memory is lacking. Furthermore, the neural mechanisms underlying such geometric reconfigurations remain unknown, particularly the role of oscillatory and aperiodic neural activity. Here, we addressed these questions using simultaneous magneto- and electroencephalography (M/EEG) recordings from healthy human participants performing a multi-item, multi-feature delayed match-to-sample visual WM task. Neural subspaces were estimated using dimensionality reduction techniques, and their geometry was characterized by quantifying the orthogonality between subspaces, distances between memory representations and subspace shapes. Participants maintained orientation and shape information simultaneously while task demands required either deprioritizing one feature or updating the representation of the prioritized feature. We found that subspace geometry was flexibly reshaped according to task demands: prioritization expanded representational subspaces, whereas de-prioritization shrunk them. In addition, subspaces representing different feature domains were generally oblique and became more orthogonal only under high cognitive demands. Finally, these task-dependent changes in subspace geometry were largely driven by interactions between aperiodic neural activity and slow oscillations in the delta and theta frequency bands.