Dyllan Simpson¹ (), Joseph Saito¹, Timothy Brady¹; ¹University of California San Diego

In a given context, some visual stimuli are consistently remembered while others are consistently forgotten resulting in the phenomenon referred to as stimulus memorability. However, observers’ metacognitive access to stimulus memorability is unclear, with studies providing mixed evidence for the ability to predict which images will be consistently remembered versus forgotten. Here, we hypothesized that inconsistent access to memorability may reflect reliance on a unique distribution of features that does not universally overlap with those that are most distinct. In Experiment 1, we trained separate ridge regression models to predict memorability and judgments of learning (JOLs) within the THINGS image database. Consistent with our prediction, we found that the visual features that predicted memorability and JOLs overlapped, but not perfectly. For example, bathroom-related features were positively correlated with memorability in the database but were negatively correlated with JOLs. In Experiment 2, we manipulated the prevalence of the unique and overlapping features identified by each model within a new set of images to see how this would influence the correspondence between JOLs and memorability. In doing so, we found that selectively overrepresenting the features that originally predicted high memorability eliminated their distinctiveness and inverted rank-order memorability, but did not invert rank-order JOLs. However, when we overrepresented the features that originally predicted both high memorability and JOLs, both rank orders inverted. Two additional experiments with scenes and sentences confirmed JOLs rapidly calibrated to local statistics producing correspondence. The dissociability suggests that metacognitive access to stimulus memorability arises incidentally when the features used to form JOLs happen to align with the features that drive memorability in the database. We speculate that these unique feature distributions reflect visual statistics acquired across different timescales, with memorability primarily reflecting the flexible statistics of the database and JOLs primarily reflecting the stable statistics of the natural environment.