Uncertainty-Aware Deep Learning Framework for Sex Attribution in Prehistoric Hand Stencils

A recent preprint on arXiv presents a deep learning framework that accounts for uncertainty in identifying the biological sex of individuals who created hand stencils during the Upper Paleolithic era. The research, found at arXiv:2608.14539, tackles issues like the absence of definitive ground truth, differences between modern and ancient populations, and image quality degradation. Conventional morphometric techniques struggle due to significant structural overlap among sexes, limited applicability across populations, and subjective feature selection. The new framework effectively models, propagates, and combines uncertainty throughout the analysis. It integrates dual image processing, contour extraction, silhouette augmentation, diverse model architecture, and ensemble decision-making, producing twelve potential silhouette representations per stencil to address boundary uncertainties. This study is significant for cultural heritage and digital archaeology, enhancing sex attribution in prehistoric art.