This paper presents a metric learning-based deep convolutional neural network (CNN) applied to an archaeological dataset. The model, trained on thousands of artifact images spanning 1.4 million years of Levantine history, achieved accuracy comparable to archaeologists in identifying artifacts by site and period. Blind validation experiments showed the model outperformed archaeologists outside their areas of expertise. A community detection algorithm, applied to the model's confusion matrix, revealed meaningful affiliations across sites, as demonstrated in a Natufian case study. The model's potential to uncover new patterns in archaeological data is highlighted.

Abraham Resler, Reuven Yeshurun, Filipe Natalio, Raja Giryes