Decoding the Past: An Uncertainty-Aware Deep Learning Framework for Sex Attribution in Prehistoric Hand Stencils
A study presents an uncertainty-aware deep learning framework for sex attribution in prehistoric hand stencils. The framework uses dual image processing, dual contour extraction, structured silhouette augmentation, model architectural diversity, and ensemble-based decision aggregation. It generates twelve plausible silhouette realizations per stencil and processes them with two ensembles of ten deep neural networks each (EfficientNet-B3 and MobileViT-S) trained on 14,036 contemporary hand samples. A triangulated validation scheme integrates ensemble predictions with unsupervised 2D latent representations.
Researchers developed a deep learning framework to determine the biological sex of individuals who created Upper Paleolithic hand stencils. The approach addresses challenges such as lack of ground truth, population differences, and image degradation by explicitly modeling uncertainty. The pipeline generates multiple silhouette realizations per stencil and uses ensembles of EfficientNet-B3 and MobileViT-S models trained on contemporary hand data, with a triangulated validation scheme.
The framework's uncertainty-aware design, combining dual contour extraction and ensemble diversity, aims to mitigate boundary uncertainty in degraded prehistoric images. The use of two distinct architectures (EfficientNet-B3 and MobileViT-S) and twelve silhouette realizations per stencil suggests a focus on robustness and calibration. The triangulated validation with unsupervised 2D latent representations may help assess prediction reliability without ground truth.
This research demonstrates an application of deep learning in archaeology and anthropology, potentially enabling more objective analysis of prehistoric artifacts. The uncertainty-aware approach could be relevant for other domains with noisy or incomplete data, such as medical imaging or forensic analysis. The reliance on contemporary hand samples for training highlights the challenge of domain shift in historical data.
The framework could be offered as a tool for archaeological research institutions or cultural heritage organizations. It may also inspire similar uncertainty-aware AI solutions in fields requiring analysis of degraded or limited data. However, commercial viability depends on validation and adoption by the scientific community.
Next signals include peer review or publication in a journal, release of code or datasets, and application of the framework to other prehistoric art forms or archaeological datasets. Further validation on known-sex historical samples or collaboration with archaeologists could strengthen the method's credibility. Potential extensions may involve incorporating additional modalities or refining uncertainty quantification.