Charlotte Atzert¹ (), Filip Děchtěrenko², Jiří Lukavský², Niko A. Busch¹; ¹University of Münster, Germany, ²Czech Academy of Sciences, Prague, Czech Republic

Why are some events remembered for a long time, while others are quickly forgotten? Specifically, which image properties make certain scene images more consistently memorable than others? While the memorability of any event can be influenced by many different factors, it has long been recognized that unusual events are remembered better than typical events. Here, we investigated whether image memorability is linked to image typicality. Specifically, we tested the hypothesis that untypical images would be harder to categorize but easier to remember. We collected a dataset of scene images from several indoor scene categories with ratings of subjective typicality. For each scene category, we also compiled two sets of feature dimensions: descriptors used by human participants to discriminate between scene exemplars (e.g., number of windows, coziness) and deep neural network features based on AlexNet layer fc7. Both sets of features were strongly predictive of scene typicality ratings, indicating that scene typicality reflects the configuration of image features of a given exemplar relative to other category exemplars. We then presented typical and untypical scene exemplars to a new sample of 90 participants. In the categorization task, participants decided as quickly as possible whether a given image was an exemplar of a specific category. In the subsequent memory task, we repeated all images from the categorization task along with new foils, and participants discriminated between old and new images. We found that typical images were categorized as exemplars of their category faster and more accurately than untypical images. Conversely, typical images were more difficult to remember than untypical images. Together, these findings highlight a trade-off between scene memorability and categorizability, suggesting that both depend on how distinct a scene's features are compared to other category exemplars.

Acknowledgements: This research was supported by the German Research Foundation (BU2400/14-1) and the Czech Science Foundation (GF24-11506K).