Perceptual similarity mostly ignores within-category feature distributions: Evidence from computational modeling of human categorizations.

Categorization is a fundamental human skill that involves assigning objects to categories based on their features. Within each category, object features can vary and correlate, giving rise to within-category feature distributions. Processing such distributional information can, in principle, improve category learning. However, the extent to which people actually make use of within-category feature distributions during categorization remains unclear due to mixed empirical evidence. To investigate this question, we conducted two category learning and transfer experiments in which the variances and correlations of features within categories were manipulated based on simulation-based optimal experimental design. Using cognitive modeling within an exemplar-similarity framework, we compared a cognitive process that ignores within-category feature distributions (Euclidean similarity) with one that considers them (Mahalanobis similarity). Across both experiments (both Ns = 43), most participants' transfer phase behavior was best captured by the simpler Euclidean model, indicating that they largely disregarded within-category feature distributions. Nevertheless, a minority of participants categorized the test stimuli in line with the Mahalanobis model, albeit less consistently, suggesting that they considered distributional information to some extent. Overall, the findings indicate that similarity-based categorization processes are generally insensitive to the statistical distribution from which objects are drawn, possibly because estimating such distributions imposes substantial computational costs. (PsycInfo Database Record (c) 2026 APA, all rights reserved).