Abstract
Why are some object concepts (e.g., windshield vs. toothpaste) more memorable than others? Prior studies have suggested that visual and semantic features (e.g., color) and typicality (e.g., for birds: robin vs. penguin) of object images influence the likelihood of being remembered with mixed success (Kramer et al., 2022). One reason for these modest effects may be due to the visual and memory spaces being predominantly modeled using Euclidean geometry, which may not reflect the true structure of the space. In this study, we examined whether an entirely different geometric relationship – such as a continuous hyperbolic space which approximates discrete hierarchies – explains differences in their memorability. Specifically, we hypothesized that image concepts would be geometrically arranged in hierarchical structures and that memorability would be explained by a concept’s depth in these hierarchical trees (where deeper concepts would be less remembered). To test this hypothesis, we constructed a hyperbolic representation space of object concepts (N=1,854) from the THINGS database (Hebart et al., 2019), which consists of naturalistic images of concrete objects, and a space of 49 feature dimensions (e.g., red, tall) derived from data-driven models. Using ALBATROSS (Stier et al., in prep.), a stochastic topological data analysis technique that detects underlying structures of data, we demonstrated that hyperbolic geometry efficiently captures the organization of object concepts and their memorability better than a Euclidean geometry. Specifically, we found that concepts closer to the center of the hyperbolic representational space are more prototypical and more memorable; in contrast, there was no consistent geometric organization of memorability and typicality in the Euclidean space. Taken together, we discover that concept typicality and depth in the hierarchical structure of image concepts contribute to how likely a concept is remembered across people.