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

Karel Becerra; Boris Mederos; Dean Snow; Ramón A. Mollineda

Source-linked AI summary

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

Karel Becerra, Boris Mederos, Dean Snow, Ramón A. Mollineda

arXiv:2608.14539v1Aug 14, 2026cs.CVcs.AIcs.LG

TL;DR

Assigning biological sex to prehistoric hand stencils is limited by overlapping morphology, population differences, degradation, and absent ground truth. This study uses a multi-layered uncertainty-aware silhouette and ensemble framework, producing stable predictions for some stencils while identifying others as ambiguous. The results support treating uncertainty as analyzable evidence rather than noise in archaeological interpretation.

Problem
Prehistoric hand-stencil sex attribution lacks definitive ground truth and is limited by overlapping morphology, population variation, poor generalizability, and subjective handcrafted features.
Method
The study models, propagates, and aggregates uncertainty through silhouette extraction and augmentation, ensemble classification, and post-hoc analysis.
Results
The framework produced stable predictions for several prehistoric stencils while identifying others as intrinsically ambiguous through reduced silhouette support and lower classification margins.
Takeaways & Limitations
Confidence indicators can help mitigate overinterpretation of uncertain archaeological evidence, while anatomically organized attributions support biologically plausible classifier behavior.
Takeaways & Limitations
Prehistoric stencil sex remains inferred from contemporary reference models rather than directly observed, leaving definitive ground truth unavailable.

Abstract

from arXiv · show

Determining the biological sex of the individuals who created Upper Paleolithic hand stencils remains a challenging problem due to the absence of ground truth, population differences between contemporary and prehistoric groups, and the uncertainty introduced by image degradation. Traditional morphometric methods suffer from high structural overlap across sexes, poor cross-population generalizability, and subjective feature engineering. This study presents an uncertainty-aware deep learning framework for sex attribution in prehistoric hand stencils that explicitly models, propagates, and aggregates uncertainty throughout the analytical pipeline. The methodology combines dual image processing, dual contour extraction, structured silhouette augmentation, model architectural diversity, and ensemble-based decision aggregation. The pipeline generates twelve plausible silhouette realizations per stencil to capture boundary uncertainties, which are processed by two ensembles of ten deep neural networks each (EfficientNet-B3 and MobileViT-S) trained on 14,036 contemporary hand samples. Furthermore, a triangulated validation scheme integrates ensemble predictions with unsupervised 2D latent-space manifold mapping (UMAP + k-NN) and explainable AI spatial attributions (LayerCAM) to ensure anatomical consistency. On contemporary data, ensemble models achieve strong classification performance, with accuracies exceeding 88% in older age groups. When applied to prehistoric stencils, the framework produces both sex predictions and confidence measures of internal agreement, enabling the distinction between morphologically stable and ambiguous cases. Convergence across ensemble predictions, latent-space structure, and interpretability analyses shows that uncertainty can become a measurable component of archaeological inference, enabling robust and reproducible decoding of ancient rock art.

1. Introduction

Sex attribution from prehistoric hand stencils is promising but constrained by overlapping sexual morphology, population-transfer challenges, image degradation, and uncertain stencil boundaries. This study addresses these limitations with an uncertainty-aware framework that generates plausible silhouettes and triangulates ensemble predictions with latent-space and interpretability analyses.

Problem: Traditional morphometric methods are limited by substantial hand-size overlap, subjective handcrafted features, poor cross-dataset generalization, and restricted representation of morphology.These approaches rely on measurements such as size, angles, and digit ratios, while automated machine-learning approaches only partially addressed the limitations.
Limitations: Degradation, incomplete preservation, illumination variation, surface irregularities, and ambiguous boundaries create uncertainty in segmentation and contour delineation, while 3D reconstruction and orthoprojection can propagate additional errors.These limitations persist even when photogrammetric 3D models and 2D orthophotos reduce viewpoint- and scale-related distortions.
Contribution: The framework explicitly models, propagates, and aggregates uncertainty across the deep-learning pipeline for sex attribution in prehistoric hand stencils.It extends a prior classifier-ensemble and spatial-attribution approach (Mollineda et al. 2025) by broadening uncertainty management across analysis stages.
Contribution: Structured silhouette representations generate multiple plausible contour realizations through dual annotation, morphological perturbations, and multi-channel compositions.The strategy is designed to capture annotation and morphological uncertainty rather than rely on a single extracted boundary.
Contribution: A triangulated validation scheme combines ensemble predictions, UMAP + k-NN latent-space analysis, and aggregated LayerCAM maps to provide convergent evidence and diagnose epistemic stability.Nine prehistoric hand-stencil cases are analyzed jointly through predictions, confidence measures, latent-space organization, and spatial attribution patterns, spanning highly consistent to ambiguous outcomes.

2. Related work

Research on sex attribution from prehistoric hand stencils has progressed from discriminant analyses and handcrafted measurements toward geometric morphometrics and deep-learning pipelines. However, population-transferability, absent ground truth, and uncertainty from manually delineated, degraded stencil contours remain central challenges.

Methodological evolution: The literature has shifted from linear measurements and digit ratios toward landmark-based morphometrics and automatically learned features from contemporary hand data.The reviewed studies span Paleolithic, simulated, and contemporary hand data, with feature extraction ranging from handcrafted measurements to deep learning.
Traditional discriminant methods: Snow (2006) introduced a two-stage discriminant approach using hand and finger lengths from 111 modern European adults to separate larger-handed adult males from smaller-handed individuals.The smaller-handed group could include adult females and subadults in archaeological contexts.
Automated feature extraction: Wang et al. (2010) automated contour extraction with HSV conversion and K-means clustering, then normalized geometric features—including the D2:D4 ratio—by middle-finger length.The pipeline extracted fingertip and valley landmarks, finger and palm dimensions, and other measurements from contemporary hand images.
Population generalizability: Models trained on one population can generalize poorly: Snow’s U.S.-trained discriminant functions produced heavily biased classifications on an independent French dataset because of average hand-size differences.Galeta et al. (2014) evaluated this issue using 100 French right-hand prints and leave-one-out validation.
Persistent limitations: Prehistoric-stencil inference lacks definitive ground truth, while manually traced silhouettes are vulnerable to observer variation, pigment diffusion, rock texture, and preservation, allowing contour errors to amplify prediction uncertainty.These limitations affect both the validity of contemporary reference models and the reliability of silhouette-based classification.

3. Materials and methods · 3.1. Overview

The study infers biological sex from prehistoric cave hand stencils using a silhouette-based computer vision pipeline. The pipeline comprises silhouette extraction and augmentation, ensemble-based classification, and post-hoc analysis.

3.1. Overview: The pipeline targets biological-sex inference from prehistoric cave hand stencils using silhouette-based computer vision.Its overall organization is illustrated in Figure 2.
3.1. Overview: The first phase extracts and augments stencil silhouettes before classification.
3.1. Overview: The second and third phases perform ensemble-based silhouette classification followed by post-hoc analysis.

3.2. Datasets

The study combines 14,036 annotated contemporary hand X-rays for DNN training with nine prehistoric hand stencils used as independent case studies. All samples are represented as binary hand silhouettes within a unified experimental framework.

Data representation: Both contemporary radiographs and prehistoric stencils are converted into binary hand silhouettes, enabling their integration within a unified experimental framework.This standardized representation supports the proposed methodology across both data sources.
Contemporary dataset: 14,036 left-hand X-rays from pediatric patients aged 1 month to 19 years were collected from two institutions and split into 12,611 training and 1,425 validation samples.The institutions contributed 2,983 and 11,053 images, respectively.
Contemporary dataset: The contemporary dataset spans age-related class distributions concentrated between 10 and 16 years, when sexually dimorphic hand traits become more stable and pronounced.Male and female classes show unimodal and bimodal distributions, respectively.
Prehistoric case studies: Nine prehistoric hand stencils from two sources serve as independent case studies, comprising four well-preserved El Castillo samples and five previously analyzed images from at least four sites.The El Castillo samples were selected for contrast and clear edges, while the second subset had prior sex-attribution hypotheses from Mollineda et al. 2025.

3.3. Methodology

The methodology treats uncertainty as a measurable, propagated component of sex attribution by combining multiple silhouette representations, diverse ensembles, hierarchical aggregation, and post-hoc validation. This controlled consensus system addresses degraded imagery, contour ambiguity, model variability, and limited ground truth without relying on extensive hyperparameter optimization.

3.3. Methodology: The framework propagates uncertainty through silhouette extraction, augmentation, ensemble classification, hierarchical aggregation, and post-hoc structural and interpretability validation.Its core principles are multiplicity before commitment, structured perturbation, architectural diversification, hierarchical aggregation, and explicit internal-agreement quantification.
Phase 1 Silhouette extraction and augmentation: Twelve silhouette representations per stencil combine dual image processing, independent contours, binary shape operators, morphological interpolation, and three-channel variants.The dual paths preserve alternative interpretations of degraded cave imagery, while structured perturbations represent admissible contour and morphological variability.
Phase 2 Ensemble-based classification of silhouettes: Two architectures, EfficientNet-B3 and MobileViT-S, each use 10 independent runs across 12 variants, producing 240 probabilistic predictions per stencil.Predictions are fused using sum-based and max-based strategies, with silhouette support rate measuring cross-variant morphological agreement.
Phase 3 Post-hoc analysis: Post-hoc validation compares ensemble predictions with 2D manifold k-NN classification and aggregates 120 EfficientNet attribution maps using the geometric median.Agreement in latent space tests structural robustness, while aggregated maps reduce explanation noise and model-specific artifacts.
Phase 2 Ensemble-based classification of silhouettes: The consensus strategy aggregates weak predictions across silhouette variants, architectures, and model instances to improve robustness to outliers and stabilize classification outcomes.This diversity addresses uncertainty from absent ground truth, domain shifts, and subjective prehistoric contour tracing.
3.3. Methodology: Explicit hyperparameter optimization was omitted to prioritize an accessible inference framework, leaving more thorough tuning as a stated avenue for improvement.The methodology therefore emphasizes usability for non-specialists rather than maximizing predictive performance through comprehensive benchmarking.

4. Experiments

Experiments show that ensemble predictions are accurate on contemporary silhouettes and stable across architectures, while confidence metrics and independent latent-space analyses quantify agreement and ambiguity in prehistoric cases. Model relevance concentrates on anatomically meaningful finger geometry and interdigital spacing rather than global hand or background structure.

Contemporary benchmark: Contemporary classification accuracy improved with subject age, likely reflecting denser older-age sampling and increasingly expressed sexual dimorphism.Performance was strongest in the older age strata, while the [0–6) cohort was comparatively less supported.
Contemporary benchmark: 76.2% ensemble accuracy exceeded 68.1% ± 1.9% for individual EfficientNet-B3 instances in the [0–6) age stratum.Ensembles consistently outperformed mean individual-instance performance, especially among the youngest subjects.
Contemporary benchmark: EfficientNet-B3 generally outperformed MobileViT-S, although the restricted task and absence of hyperparameter tuning preclude broad architectural-superiority claims.The architecture comparison covered specific model variants applied to a single specialized task.
Prehistoric attribution: High SSR and ACM identified morphologically stable prehistoric cases, whereas lower values corresponded to ambiguous stencils and preserved uncertainty rather than forcing confident classifications.Aggregation across 120 individual predictions per stencil produced stable agreement across architectures and aggregation strategies.
Latent-space validation: In all nine cases, k-NN labels matched EfficientNet-B3 ensemble predictions, showing that sex-related structure remained separable after 2D dimensionality reduction.Support was strongest for img_26, img_28, Cosquer, and El Castillo 25, while moderate support did not change assignments.
Explainability: LayerCAM relevance concentrated on fingers, interdigital spaces, phalanges, and metacarpal heads, with minimal activation in the palm and wrist.The spatial pattern indicates reliance on finger geometry and relative spacing rather than global hand size or background structure.

5. Discussion

The framework treats uncertainty as a measurable, propagated component of sex attribution, combining predictive, latent-space, and attribution evidence to distinguish stable from ambiguous classifications. Its archaeological interpretation remains limited by reliance on contemporary reference populations and the absence of prehistoric ground truth.

5. Discussion: Uncertainty is modeled and propagated across image processing, contour extraction, silhouette perturbation, representation, and model stages before hierarchical aggregation.This multi-layered design treats uncertainty as information rather than noise to suppress.
5. Discussion: UMAP–k-NN validation and geometric-median aggregation of LayerCAM maps provide structural and explanatory checks beyond probability estimates.Each LayerCAM map combines 120 attribution instances, stabilizing relevance patterns across contour variants and stochastic model variation.
5. Discussion: Attributions consistently emphasize fingers, interdigital spaces, phalanges, and metacarpal heads while assigning comparatively little relevance to the palm and wrist.These regions include structures previously identified as important sources of sexual dimorphism in skeletal studies.
5. Discussion: Agreement among predictive aggregation, latent-space separability, and attribution coherence supports morphologically stable classifications, whereas divergence identifies structurally ambiguous cases.Cases with SSR ≈1 show sharply localized, organized activation patterns; disagreement or reduced silhouette support produces more diffuse relevance.
5. Discussion: The framework supports archaeological interpretation through joint analysis of predictions, confidence, latent-space organization, and attribution patterns rather than binary sex labels alone.Strongly consistent cases are characterized by high silhouette support, large classification margins, consistent latent-manifold positioning, and agreement between ensemble and k-
5. Discussion: Because prehistoric ground-truth sex labels do not exist, the method relies on contemporary reference populations and assumes some sexual-dimorphism signatures persist across long temporal scales.This assumption cannot be directly verified for Upper Paleolithic populations, motivating richer plausible-shape representations, predictive uncertainty modeling, and broader probabilistic scaling.

6. Conclusions · Ethics approval

The study presents uncertainty as an informative component of sex attribution in prehistoric hand stencils, modeling, propagating, and aggregating it across the analytical pipeline. Convergent evidence supports morphologically stable classifications, whereas disagreement identifies ambiguous cases requiring cautious interpretation.

6. Conclusions: The framework models, propagates, and aggregates uncertainty across image acquisition, annotation, silhouette representation, model training, and decision making.It combines controlled variability with hierarchical aggregation.
6. Conclusions: Contemporary hand-silhouette experiments show that ensemble deep learning reliably captures sexually dimorphic morphology, especially in age groups where these traits are more strongly expressed.
6. Conclusions: Transferred to prehistoric stencils, the framework yields stable predictions for some cases while identifying others as intrinsically ambiguous through reduced silhouette support and lower classification margins.
6. Conclusions: Triangulation combines ensemble classification, unsupervised interpretable 2D latent-space validation, and aggregated LayerCAM explanations.Agreement among these perspectives supports consistent latent representations and anatomically meaningful visual cues.
6. Conclusions: Agreement across predictive, structural, and interpretability analyses provides convergent evidence for morphologically stable classifications.
6. Conclusions: Divergence across analyses highlights cases where the available evidence remains insufficient for definitive attribution, making uncertainty a reliability signal rather than merely an error source.
6. Conclusions: Future work should develop more explicitly probabilistic formulations to represent input variability and quantify prediction uncertainty.Such developments may help distinguish morphological ambiguity from model-related uncertainty.