Waragon Phusuwan^1,2 (), Xue-Xin Wei³, Chaipat Chunharas^1,4,5; ¹Cognitive Clinical and Computational Neuroscience Center of Excellence, Chulalongkorn University, Bangkok, Thailand, ²Medical Sciences, Faculty of Medicine, Chulalongkorn University, Bangkok, Thailand, ³Department of Neuroscience, Department of Psychology, Center for Perceptual Systems, Center for Learning and Memory, Center for Theoretical and Computational Neuroscience, The University of Texas at Austin, Austin, Texas, U.S.A., ⁴Division of Neurology, Department of Medicine, Faculty of Medicine, King Chulalongkorn Memorial Hospital, Thai Red Cross Society, Bangkok, Thailand, ⁵Chula Neuroscience Center, King Chulalongkorn Memorial Hospital, Thai Red Cross Society, Bangkok, Thailand

Visual working memory is often conceptualized as the persistence of sensory-like representation. Consistent with this view, decoding analysis in early visual cortex reveals cross-generalization between perception and memory, suggesting that their representations may be shared. However, quantifying this alignment directly from the voxel-level activity has been challenging, as noise and task-irrelevant variance can obscure underlying geometry. To address these limitations, we develop an approach that analyzes the geometry of the task-relevant subspaces inferred from the Inverted Encoding Model (IEM). We applied our approach to analyze a published fMRI dataset collected while six participants performed a visual localizer task (perception) and a delayed-orientation recall task (WM). IEM models were fitted to both task conditions to obtain voxel-to-IEM channel loadings as an effective coordinate basis. While previous work on IEM generally focused on the response of individual channels, our approach instead analyzes the geometrical properties of the subspaces defined by these IEM channels. We evaluate the subspace geometry using Canonical Correlation Analysis (CCA) to identify the principal angles, and Basis Procrustes Analysis to calculate the principal rotation angle between the coordinate axes of perception and WM feature subspaces. In contrast to the raw BOLD signal, where subspaces for perception and working memory appeared orthogonal, the feature-specific latent subspaces revealed significant alignment. Moreover, CCA and Basis Procrustes methods showed stable shared geometry from V1 through IPS0. Interestingly, while alignment of the IEM axes remains stable in early visual cortex, it increases slightly in higher-order parietal regions (IPS1). Our analysis of task-relevant subspaces revealed substantial alignment between perception and working memory representation – an effect that appears to be preserved across visual hierarchies. The alignment of the subspaces may inform computations during tasks that involve both perception and working memory.

Acknowledgements: This research project is supported by the Second Century Fund (C2F), Chulalongkorn University.