The visual memorability of natural warning patterns: insights from humans and machines

While some animals camouflage themselves, others advertise that they are toxic using bright colours and salient stripes and/or spots (‘warning patterns’). Their striking appearance is thought to warn off predators: a memorable pattern may help predators learn about toxicity and discourage future attacks on similar prey. However, how warning patterns influence visual memory has never been documented. Research suggests that when glancing at a picture, people do not intuitively know what makes it memorable or forgettable, but they remember and forget the same images (i.e., there is high inter-subject consistency). This means that the likelihood of remembering a picture (its ‘memorability’) can be computationally predicted from the visual information contained in the picture. Memorable images also lead to stronger neural firing when processed by real and artificial visual systems. We used a database of Lepidoptera (butterfly/moth) images, some of which carry warning signals (aposematic: AP), and some which do not (non aposematic: nAP). We measured human memorability for both AP and nAP Lepidoptera and examined the sources of memorability variation across subjects. Observers studied images while providing subjective ratings (1-10) of memorability, followed by a recognition test (‘Seen before?’). Memorability was computed as the proportion of subjects who remembered previously seeing each image. AP species appeared subjectively more memorable than nAP ones, but, on average, they were not better remembered. Remarkably, AP species led to high inter-subject consistency in memorability (Spearman’s rho = .79), but consistency for nAP species was comparatively low (Spearman’s rho = .37). When we exposed our Lepidoptera patterns to deep neural networks trained for object classification, we found that AP species that were memorable to humans also evoked stronger neural responses in some hidden layers. Taken together, these findings suggest that warning patterns might exploit shared visual mechanisms that underlie successes and failures in picture recognition.