Federico De Filippi¹ (), Olivier Penacchio^(1,2), Akira R. O'Connor¹, Julie M. Harris¹; ¹University of St Andrews, St Andrews, United Kingdom, ²Computer Vision Center, Universitat Autònoma de Barcelona, Barcelona, Spain

What makes certain images stick in our memory more easily than others? While memorability has been linked to high-level visual processing, the role of low-level vision is still not fully understood. In previous work, we explored vision and memory for the striking ‘warning patterns’ that toxic butterflies display to deter predators. Warning patterns evoked stronger and less sparse activity when exposed to a biologically inspired computational model of low-level vision, setting them apart from non-toxic species and other natural images. Here, we used our vision model to measure magnitude and sparseness of responses to natural and man-made textures from the Oxford Describable Textures Dataset. Using the same statistics that distinguish toxic from non-toxic butterflies, we selected sets of textures that evoked high vs. low activity and presented them to humans in a memory test. Observers (N = 100) viewed textures and reported on a 10-point scale their impression of how ‘memorable’ they appeared. Next, observers performed an old/new recognition test (‘Seen before?’). We found that textures that evoked high neural activity were rated as more memorable (mean rating ± SE: high = 0.50 ± 0.03, low = 0.37 ± 0.03, p < .001) and were also more easily remembered (mean d’ ± SE: high = 2.13 ± 0.19, low = 1.51 ± 0.16, p < .001) than those that evoked low activity. We compared human performance to a state-of-the-art deep learning model for memorability prediction, ResMem (Needell & Bainbridge, 2022, Comp Brain & Behav. DOI:10.1007/s42113-022-00126-5). The model was moderately predictive of both human ratings (Spearman’s rho = 0.46, p < .001) and proportion correctly remembered (Speaman’s rho = 0.38, p < .001). Our findings suggest that early visual computations, which distinguish toxic from non-toxic animals in nature, may also differentiate memorable images from forgettable ones in humans.

Acknowledgements: This work is funded by the Biotechnology and Biological Sciences Research Council (United Kingdom Research and Innovation)