Abstract
Why do we remember some things while forgetting others? Prior work has demonstrated remarkable consistency in what stimuli people will later remember, a quantifiable and robust property known as memorability. To explain memorability of faces and scenes, prior research has used singular image features or a linear combination of image features with limited success. In order to provide a richer account of memorability, we utilize a spatial framework with individual images represented as points in a multidimensional space generated from a broad, general stimulus set. Specifically, we leveraged THINGS, a naturalistic object image database of 26,107 images that representatively samples concrete objects to examine the image features that most strongly influence memorability and whether it is the most prototypical or most atypical items that are most memorable. We focus on the role of semantic and visual features and their distribution in determining which images are best remembered. We collected memory performance data from 13,946 participants and utilized three complementary measures of typicality including human typicality judgments, similarity across object space dimensions, and across deep neural network features to relate stimulus typicality to memorability. Our results suggest that semantic information has a stronger influence on memorability than visual information with a slight bias toward the most prototypical items being the most memorable. These findings run counter to the predominant view that semantic information is not required to explain what is memorable and that the most atypical images are best remembered. Our findings shed new light on the determinants of what makes something memorable that could only be found using a large, representative dataset.