Ivette Colón¹, Timothy Rogers¹; ¹University of Wisconsin - Madison

Neural representations of visual categories are reliably distributed across cortex, yet it remains unclear how much of their internal structure is shared across individuals versus individually specific, and how stable such structure is over time (Haxby et al., 2014). Specifically, do neural representations preserve a shared categorical organization while exhibiting stable, individual-specific representations across time? Here, participants viewed controlled sets of familiar face, place, and object images during two fMRI sessions separated by several weeks. Neural responses were analyzed using sparse multiclass decoding to assess category information across cortical activity patterns. Mean within-session decoding accuracy across participants was 70.2% in Session 1 and 66.9% in Session 2 (chance = 33%). To quantify the stability and individuality of representational structure, we trained decoding models within individuals and evaluated their performance across sessions and across participants. To enable cross-session and cross-participant comparisons despite differences in voxel sampling, we aligned voxel sets using bipartite graph matching based on spatial proximity in standardized brain space. Within individuals, cross-session decoding was highly reliable: models trained on Session 1 generalized to Session 2 with a mean accuracy of 65.5%. In contrast, decoding models trained on anatomically-matched data from other participants performed substantially worse when applied to a given individual (42.2%). The within-individual advantage was large when expressed relative to the between-participant distribution (mean z-score = 3.18), demonstrating pronounced individual specificity in representational organization. Importantly, above-chance decoding accuracy persisted when analyses were restricted to voxels outside canonical category-selective regions, indicating that individual-specific semantic structure is distributed broadly across cortex rather than confined to classical functional localizers (Julian et al., 2012). Together, these results show that joint representations of faces, places, and objects are stable over time within individuals, reliably distinct across people, and supported by distributed neural patterns.