Hyewon Willow Han^1,2, Yalda Mohsenzadeh^1,2; ¹Western University, ²Vector Institute for Artificial Intelligence

What makes some images of the same object concept more memorable than others? Recent findings suggest that semantic features exert a stronger influence on the memorability of object images than perceptual features (Kramer et al., 2023). However, the factors driving variability in image memorability within the same object concept remain elusive. Here, we address this question through the lens of concept manifold geometry (Sorscher et al., 2022). Using a large-scale naturalistic object image collection with diverse concepts (Hebart et al., 2019), we first identify the representational space that best predicts memorability by linearly probing layers of deep neural network (DNN) vision models. We then characterize, for each object concept, the geometry of memorable and forgettable submanifolds within the overall concept manifold. Our analysis reveals that memorable exemplars occupy higher-dimensional submanifolds than forgettable exemplars, and that the distance between the submanifolds increases as the overall concept manifold has a larger radius but lower dimensionality. Furthermore, after accounting for between-concept variability, semantic and visual dimensions (Hebart et al., 2023; Contier et al., 2024), together with local manifold geometry, all make significant contributions to predicting image memorability. Lastly, by estimating within each concept the direction that best separates memorable from forgettable images, we show that these memorability directions are largely concept-specific and that variation in local dimensionality and radius along them shapes which images are remembered or forgotten. Together, these results suggest that image memorability within an object concept is governed by concept-specific manifold geometry rather than by a single global set of features.