A deep learning approach to detect and visualise sexual dimorphism in monomorphic species

Abstract

Sex recognition is facilitated by dimorphism in some traits. However, humans often fail to find the traits that allow to distinguish between sexes in other species. Deep learning has the potential to surpass humans in identifying cryptic differences between sexes, but, so far, has rarely been used to assess sexual dimorphism. In this study, we evaluated (i) the ability of a fine-tuned classification neural network, EfficientNet, to find differences between sexes in a species that appears monomorphic to humans, the sociable weaver (Philetairus socius). We then assessed (ii) the benefits of Grad-CAM visualisation techniques to understand which parts of the individuals are used by the network to differentiate the sexes. We trained 10-folds cross-validation models on more than 4,500 pictures of the head from more than 1,300 individuals. Our results show that the network can predict sex of sociable weavers with an accuracy of 76%, which is considerably higher than humans performance, and that the model was similarly good at predicting females and males. When interpreting the probability of being classified to one sex, our results further reveal an effect of the interaction of sex with age on the confidence score of the models which shows that younger males are less masculine than older ones, and older females more masculine than younger ones. Finally, using Grad-CAM we found that the model mostly uses the beak region to predict the sex of individuals. Overall, this work shows that artificial intelligence has the potential to be a non-invasive sexing tool, surpassing human capabilities and aiding in pinpointing potential cryptic dimorphic body parts that have yet to be identified. In birds, half of the worlds species appear sexually monomorphic to humans, and re-evaluation of species dimorphism with this type of methods could deepen our understanding of the effect of selection on animal traits.