Chihye Han¹ (), Michael Bonner¹; ¹Johns Hopkins University

Recent work in neuroscience challenges the traditional view that visual cortex reduces sensory inputs to low-dimensional representations, suggesting instead that neural population codes operate on high-dimensional manifolds. While previous work has demonstrated reliable individual differences in low-dimensional cortical responses, the structure of individual variation across higher-dimensional spaces remains poorly understood. Here, we investigated whether individual differences in visual experience are encoded in high-dimensional cortical response patterns during naturalistic movie viewing. We first replicated and extended previous work by showing that the reliable variance in movie-evoked fMRI responses is distributed across the full spectrum of latent dimensions in cortical activity, following a power-law distribution over latent-dimension ranks (here detected across hundreds of orthogonal dimensions). These effects were observed throughout visual cortex, including occipital, ventral temporal, and lateral temporal regions as well as in higher level semantic regions, including the supramarginal and angular gyri. We next constructed individual difference matrices (IDMs) to characterize intersubject variability in these high-dimensional representations. Our IDM analysis revealed reliable individual differences across multiple orders of magnitude of latent-dimension ranks. These high-dimensional differences persisted after accounting for voxelwise inter-subject correlations, indicating that they reveal aspects of individual variability that cannot be detected with conventional analysis methods. Further, we found that the IDMs vary across latent-dimension ranks, suggesting that distinct patterns of individual variability can be observed when considering different subsets of latent dimensions. Remarkably, we found that these high-dimensional differences in cortical representations correlate with differences in how subjects verbally described their visual experiences, quantified through semantic embedding distances between recall transcripts. Together, these results reveal a rich landscape of individual differences in visual processing that extends far beyond previously studied low-dimensional representations, suggesting that variation in subjective visual experience is encoded within largely unexplored high-dimensional neural structure.