Decoding the Past: An Uncertainty-Aware Deep Learning Framework for Sex Attribution in Prehistoric Hand Stencils

Uncertainty quantification is critical for archaeological AI: The framework explicitly models multiple sources of uncertainty—including image degradation, boundary ambiguity, and population differences—rather than treating predictions as definitive, which is essential when analyzing 40,000-year-old artifacts where ground truth doesn't exist.

Deep learning can achieve 88%+ accuracy on contemporary hands but faces real limitations with prehistoric data: While the ensemble models performed strongly on modern hand samples (>88% accuracy in older age groups), the paper acknowledges that applying these models to prehistoric stencils involves substantial extrapolation across time periods and populations, highlighting the gap between lab performance and real-world archaeological application.

Multiple validation approaches strengthen confidence in uncertain predictions: The study uses three complementary methods—ensemble predictions, unsupervised manifold mapping, and explainable AI visualization—to cross-check results and ensure anatomical consistency, demonstrating that responsible AI for archaeology requires triangulation rather than relying on a single model's output.

Structured uncertainty representation improves interpretability: By generating twelve plausible silhouette variations per stencil and tracking how predictions vary across these variations, the framework makes uncertainty visible and measurable to archaeologists, rather than hiding it behind a single confidence score.

Population differences remain a fundamental challenge: The paper identifies that contemporary and prehistoric populations have different hand morphologies, meaning even well-trained models may not reliably transfer to archaeological contexts—a reminder that AI tools require domain expertise and cannot replace careful archaeological judgment.