Deep learning classification of tooth scores made by different carnivores: achieving high accuracy when comparing African carnivore taxa and testing the hominin shift in the balance of power

Abstract

Listen

Translate article

The balance of power (that is the dominance on the predation arena between carnivore competitors and hominins) remains controversial. One reflection of this is the carnivore modification of hominin bones. During human evolution, hominins were first prey and then predators of other animals, including carnivores. Modifications reported on some hominin bones could result from primary predators feeding on them or post-depositional modifications by scavengers. Determining carnivore agency would be crucial to interpret such information. Here, a series of computer vision models based on convolutional neural networks is presented, comparing five different types of carnivores jointly and then pairwise. It is shown how such models contain different heuristics regarding specific carnivore taxa, which regarding tooth marks made by lions and spotted hyenas can be accurately classified by as much as 92% of the testing set. The present study also shows the potential of transfer knowledge in building accurate classification of images and for taphonomic interpretation. The application to tooth marking on a 500 ka hominin femoral shaft indicates that by that time, carnivore modifications of human remains may have resulted from post-depositional scavenging rather than by predation.