Deep learning identification of anthropogenic modifications on a carnivore remain suggests use of hyena pelts by Neanderthals in the Navalmaíllo rock shelter (Pinilla del Valle, Spain)

Abstract

The identification of anthropogenically-modified carnivoran bones in archaeological sites is rare in Pleistocene contexts, especially in the most ancient periods. Neanderthal groups have clearly shown a great variety of subsistence activities and the use of carnivoran resources, until rare, is also present in some archaeological sites.

However, the identification of the bone surface modifications (BSM) that allow us to infer the presence of anthropogenic marks in faunal remains are usually difficult to be differentiate among other BSM. Recently, several statistical and computing techniques have been developed to differentiate among different types of BSM in an objective way.

To date, the most powerful approach is the use of Convolutional Neural Networks, which are the essential part of what is referred to as Deep Learning. In this work, ResNet50 and Inception V3 models are used through transfer learning. The algorithm architecture reaches an accuracy of >96.3% when differentiating among experimental trampling, cut and tooth marks. Once the transfer models were re-trained with the experimental BSM, they were used to classify several archaeological BSM previously identified as cut marks by human analysts. These BSM have been found on a bear ulna and on a hyena phalanx, both recovered at the Navalmaíllo Rock Shelter (Madrid, Spain).

The BSM located on the hyaena phalanx have been identified as cut marks with a high probability while marks on the bear ulna are non-anthropogenic. This bone adds to the existing sample of anthropogenically-modified carnivoran elements by Neanderthal populations and hint to use of carnivore pelts by Neanderthals.