Predictive processing theories propose that efficient real-world vision is achieved by comparing incoming sensory data with predictions generated from internal models. Given their pivotal role in perception, can individual differences in these models account for idiosyncrasies in perception?
We used an inter-subject representational similarity analysis framework to explore whether idiosyncratic variation in scene perception and gaze behavior can be predicted by similarities in individuals’ internal models of the world. To characterize these internal models, participants drew prototypical versions of indoor scenes. Inter-subject representational dissimilarity matrices (IS-RDMs) were constructed by quantifying similarities in these drawings across all pairs of participants, utilizing feature maps extracted from a deep neural network.
Behavioral IS-RDMs were constructed capturing inter-subject similarities across various performance metrics: (i) a categorization task involving indoor scene photographs, (ii) five subjective ratings (typicality, familiarity, attractiveness, usability, and complexity) of the same photographs, and (iii) gaze behavior in a free-viewing paradigm. Comparing IS-RDMs from the drawings with behavioral IS-RDMs revealed shared inter-subject similarities between participants’ drawings (and thus their internal models) and (i) categorization efficiency, as well as (ii) ratings of typicality, usability, and complexity. A preregistered replication confirmed the link between shared internal models and shared categorization performance. However, no significant correlation was observed between IS-RDMs derived from participants’ typical drawings and IS-RDMs based on inter-subject similarities in the fixation counts or gaze dwell times on objects.
Overall, we demonstrated a robust link between internal models and variation in perception, but more work is needed to establish connections to scene exploration.