PaperlayerView exact revision on arXiv

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:2608.14539v1cs.CVcs.AIcs.LG

TL;DR

Inferring the sex of prehistoric hand-stencil makers is challenging because hand measurements overlap across sexes and ages and generalize poorly across populations. This paper propagates uncertainty through silhouette preparation and ensemble classification, then hierarchically aggregates complementary observations. It reports stable predictions for several prehistoric stencils while identifying others as ambiguous.

  • Problem

    Inferring the sex of Paleolithic cave artists is challenging because traditional hand measurements overlap across sexes and ages and generalize poorly across populations.

    Source: §1 · p. 2

  • Method

    The framework propagates uncertainty through silhouette preparation and ensemble classification, using hierarchical aggregation of complementary observations to infer sex from prehistoric hand stencils.

    Source: §3.1 · p. 7, §6 · p. 25

  • Results

    The framework produced stable sex predictions for several prehistoric hand stencils while identifying other cases as ambiguous through reduced silhouette support and lower classification margins.

    Source: §6 · p. 25

  • Takeaways & Limitations

    Confidence indicators can guide interpretation of prehistoric hand stencils by helping mitigate overinterpretation of uncertain archaeological evidence.

    Source: §6 · p. 25

    Prehistoric sex attributions lack definitive ground truth and rely on contemporary reference models rather than direct observation.

    Source: §2 · p. 6

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

The study proposes a deep-learning framework that explicitly models, propagates, and aggregates uncertainty in sex attribution from prehistoric hand stencils. It combines multiple plausible silhouette representations with ensemble predictions, latent-space analysis, and spatial interpretability, demonstrating its application through nine nonrepresentative case studies.

1. Introduction

  • 60–65% accuracy often characterizes traditional morphometric methods, alongside sex-and-age overlap and poor population generalizability (Snow 2006; Snow 2013; Jowaheer and Agnihotri 2011; Galeta et al. 2014).Same-hand stencil measurements can vary by up to 5 mm. Automated approaches still depend on human-designed feature selection, leaving abstract morphological patterns unexplored (Wang et al. 2010; Fernández-Navarro, Fidalgo-Casares, et al. 2025).Source: §1 · p. 2
  • Contemporary-to-prehistoric deep-learning transfer assumes that sex-informative hand morphology remains sufficiently invariant across age and long-term evolutionary change.Recent approaches train on contemporary silhouettes because annotated ancient data are unavailable, learning hierarchical image representations without manual feature engineering (Mollineda et al. 2025; Fernández-Navarro, González-Marfil, et al. 2025).Source: §1 · pp. 2–3
  • Degradation, ambiguous boundaries, geometric distortions, and absent biological ground truth require uncertainty-aware inference rather than reliance on geometric preprocessing alone.Photogrammetry and orthophotos reduce viewpoint and scale distortions but can introduce reconstruction and projection errors, particularly on irregular surfaces. Small geometric changes can affect morphometric descriptors and subsequent classification.Source: §1 · p. 3
  • The proposed framework models, propagates, and aggregates uncertainty across image processing, contour extraction, silhouette generation, and final classification.Extending the ensemble framework of (Mollineda et al. 2025), it combines dual image processing and dual contour extraction with morphological perturbations and multi-channel silhouette compositions.Source: §1 · pp. 3–4
  • Nine hand-stencil images demonstrate the framework across consistent and ambiguous predictions, without representing Upper Palaeolithic stencil morphological variability.Each case is interpreted jointly through sex predictions, confidence measures, latent-space organization, and spatial attribution patterns.Source: §1 · p. 3
  • Triangulated validation combines ensemble predictions, UMAP + k-NN latent-space analysis, and aggregated LayerCAM maps to assess convergent evidence and epistemic stability.These complementary analyses support interpretation through agreement across predictions, spatial attributions, latent-space structure, and uncertainty measures.Source: §1 · pp. 3–4

2. Related work

This review compares hand-stencil sex-attribution studies by data type, feature extraction, decision rule, and experimental protocol. It traces a shift from handcrafted morphometrics toward learned silhouette features while highlighting missing prehistoric ground truth, population differences, and segmentation uncertainty.

2. Related work

  • used sequential discriminant analyses of hand dimensions and digit ratios to distinguish adult males, adult females, and subadult males.Functions learned from 111 contemporary European-descent adults were applied to six Paleolithic stencils. Hand dimensions first separated adult males from smaller hands; D2:D4 and D2:D5 ratios then addressed the overlap between adult females and subadult males.Source: §2 · p. 4
  • automated contour extraction and handcrafted morphometry, applying an SVM trained on contemporary hands to prehistoric French handprints.HSV conversion and K-means clustering isolated contours, with fingertips and valleys supporting geometric measurements. The 33-dimensional feature vectors normalized measurements by middle-finger length; the normalized middle finger and unreliably measured thumb were excluded.Source: §2 · pp. 4–5
  • found that U.S.-trained discriminant functions generalized poorly to French hands, with population differences in hand size producing heavily biased classifications.The study compared locally fitted functions using leave-one-out cross-validation against functions from on 100 French right-hand prints. The authors concluded that modern morphometric references were unreliable for estimating prehistoric artists’ sex.Source: §2 · p. 5
  • Landmark-based geometric morphometrics combined shape alignment, dimensionality reduction, and discriminant classification in (Nelson, Hall, et al. 2017) and (Fernández-Navarro, Fidalgo-Casares, et al. 2025).The former used 19 landmarks from 132 contemporary stencils and cross-validation without an independent test set. The latter used 32 landmarks across contemporary, experimental, and Paleolithic samples, separating Procrustes shape coordinates from centroid size.Source: §2 · p. 6
  • (Fernández-Navarro, González-Marfil, et al. 2025) selected EfficientNetV2-S as its most stable silhouette classifier, but segmentation variability and female-biased subadult classification remained limitations.Evaluation used ten independent train-validation-test splits before predicting 45 archaeological stencils; combining scanned and experimental data helped narrow the domain gap. Alongside (Mollineda et al. 2025), this work exemplifies automatically learned silhouette features requiring robust cross-domain validation.Source: §2 · p. 6
  • Absent prehistoric ground truth, subjective contour delineation, and modern–ancient domain differences limit reliable sex attribution and motivate explicit predictive uncertainty.Pigment diffusion, rock texture, preservation, and observer variability can introduce contour errors that propagate through classification. Reliable frameworks must address segmentation variability, dataset representativeness, and domain adaptation rather than accuracy alone.Source: §2 · pp. 6–7

3. Materials and methods · 3.1. Overview

The study aims to infer biological sex from prehistoric cave hand stencils through a three-phase, silhouette-based computer vision pipeline.

3.1. Overview

  • Silhouette extraction and augmentation constitute the pipeline’s first phase for analyzing prehistoric cave hand stencils.Source: §3.1 · p. 7
  • Ensemble-based classification of silhouettes forms the pipeline’s second phase for inferring biological sex.Source: §3.1 · p. 7
  • Post-hoc analysis constitutes the third and final phase of the silhouette-based computer vision pipeline.Source: §3.1 · p. 7

3.2. Datasets

Contemporary hand radiographs supplied labeled training data, while nine prehistoric hand stencils served as case studies. Binary hand silhouettes provided a common representation for training a sex classifier and applying it to prehistoric images.

3.2. Datasets

  • Binary silhouettes unified contemporary training samples and prehistoric case studies, allowing a contemporary-trained sex classifier to generate sex-related hypotheses for ancient stencils.Source: §3.2 · p. 7

3.2.1. Contemporary data

  • 14,036 contemporary left-hand radiographs represented pediatric patients aged between 1 month and 19 years.The dataset contained 12,611 training and 1,425 validation samples. A separate Stanford test set contained 200 radiographs (100 male and 100 female), without overlap with training or validation studies.Source: §3.2.1 · p. 7
  • Clinical reports supplied sex and skeletal-age labels; silhouettes used the Segment Anything Model with zero-prompt segmentation (Mollineda et al. 2025).Source: §3.2.1 · p. 7
  • 10–16 years was the most densely represented bone-age range for both sexes; male distributions were unimodal and female distributions bimodal.The authors associate this range with more stable, pronounced sexually dimorphic hand traits. They expect broad age and morphological coverage to improve model robustness and generalization.Source: §3.2.1 · p. 7

3.2.2. Prehistoric data

  • 9 prehistoric stencils comprised four El Castillo examples and five cases previously analyzed by (Mollineda et al. 2025).The four El Castillo images offered high contrast and well-defined edges supporting reliable manual contouring. The other five came from Cosquer, Chauvet, Pech Merle, and El Castillo, enabling uncertainty-aware reassessment of earlier sex-attribution hypotheses and detailed case-by-case analysis.Source: §3.2.2 · pp. 7–8

3.3. Methodology

The methodology models, propagates, and aggregates uncertainty across silhouette extraction, ensemble classification, and post-hoc analysis. Predictions are accompanied by internal agreement measures and examined through latent-space structure and spatial attribution.

3.3.3. Phase 2: silhouette classification (sex prediction via score aggregation)

  • 240 probabilistic predictions per stencil come from processing twelve silhouette variants with ten independently initialized instances each of EfficientNet-B3 and MobileViT-S.Combining representational and model diversity is intended to stabilize consensus decisions and improve robustness to outliers.Source: §3.3.1 · p. 9, §3.3.3 · p. 13
  • Sum and max fusion produce architecture-specific predictions, accompanied by silhouette support rates measuring agreement across contour variants.Sum fusion accumulates class probabilities across models; max fusion selects each class’s highest model probability. Silhouette support is the proportion of variants consistent with the predicted class.Source: §3.3.1 · pp. 9–10, §3.3.3 · p. 13

3.3.1. Overview

  • The framework models and propagates uncertainty through multiple representations, architectural diversification, and hierarchical aggregation rather than enforcing a single deterministic interpretation.Explicit quantification of internal agreement complements structured perturbation and multiplicity before commitment.Source: §3.3.1 · pp. 8, 10

3.3.2. Phase 1: silhouette extraction and augmentation

  • Original and tonally enhanced images are independently traced, converted into cleaned binary masks, and geometrically normalized to manage image and annotation uncertainty.Background suppression and median filtering isolate hand morphology; rotation, scaling, and translation standardize poses while preserving anatomical shape.Source: §3.3.2 · pp. 10–11
  • 12 silhouette representations per stencil comprise two primary contours, four binary augmentations, and six three-channel compositions.AND, OR, averaging, and distance-based morphological mid-set interpolation generate plausible contour alternatives without compromising global anatomy. Three-channel compositions additionally vary input representation.Source: §3.3.1 · p. 9, §3.3.2 · pp. 11–12

3.3.3. Phase 2: silhouette classification (sex prediction via score aggregation)

  • Missing prehistoric labels, unresolved domain shifts, and tracing ambiguity undermine individual predictions; hyperparameter optimization was omitted to prioritize an accessible inference framework.Models train on labeled RSNA Bone Age Challenge data, while scarce high-quality prehistoric samples preclude reliable domain adaptation. Comprehensive classification benchmarking falls outside the study’s scope.Source: §3.3.3 · pp. 13–14

3.3.4. Phase 3: post-hoc analysis and interpretation

  • 1,536 EfficientNet-B3 features undergo unsupervised 2D UMAP (McInnes et al. 2018) projection, supporting complementary k-NN classification against contemporary training silhouettes.Each model receives 108 projections from twelve variants of nine prehistoric stencils. Majority voting first combines variant predictions, then consensus across model instances.Source: §3.3.4 · p. 14
  • 120 LayerCAM maps per stencil are aggregated by geometric median to suppress outliers in spatial explanations.Maps span twelve silhouette variants and ten EfficientNet-B3 instances, examining whether predictions emphasize anatomically meaningful regions such as finger proportions and palm width.Source: §3.3.4 · p. 14

4. Experiments

Ensemble predictions on nine prehistoric stencils showed strong internal agreement in several cases, while confidence measures and post-hoc analyses identified ambiguity in others. Contemporary benchmarking generally favored EfficientNet-B3 and ensemble aggregation.

4.1.1. Experimental design

  • For each architecture, ten independent model instances were trained for 100 epochs, with checkpoints selected by highest validation accuracy and independently evaluated on the test subset of contemporary hand silhouettes.The source dataset contained 14,036 pediatric left-hand X-rays spanning ages 1 month to 19 years; extracted silhouettes supplied 12,509 training, 1,416 validation, and 196 test samples, resized to 400 × 300 and normalized to [0, 1]. Random training-sample ordering was the primary source of ensemble diversity, and ensembles aggregated predictions using sum and max rules, with performance measured by overall accuracy.Source: §4.1.1 · p. 15

4.1.2. Sex prediction results on contemporary data

  • 76.2% ensemble accuracy exceeded 68.1% ± 1.9% mean individual-model accuracy for EfficientNet-B3 in the [0–6) age stratum under both aggregation rules.Ensembles consistently outperformed mean individual-model performance, and EfficientNet-B3 exceeded MobileViT-S for individual models and most ensemble configurations. This architectural comparison concerns specific variants on one specialized task without hyperparameter tuning, not general architectural superiority.Source: Table 3 · p. 16, §4.1.2 · p. 16
  • 85.4% accuracy was reached by the EfficientNet-B3 sum ensemble in the older age strata, with classification generally improving with age.The authors attribute this trend to greater representation of older children and increasingly pronounced sexual dimorphism. Confusion matrices showed more stable Male predictions and greater Female variability, especially among younger children; these differences concern class sensitivity rather than prediction confidence.Source: §4.1.2 · pp. 15–16, Table 3 · p. 16, Figure 9 · p. 17

4.2.1. Sex prediction results on prehistoric data

  • 120 predictions per architecture combined ten model instances and twelve silhouette variants, producing majority-vote stencil labels with explicit internal-agreement measures.Sum/max aggregation precedes variant majority voting; SSR measures supporting variants from 0.5 to 1, while ACM is the absolute mean signed margin, ranging from 0 to 1. EfficientNet-B3 generally produced more stable, higher-confidence predictions than MobileViT-S, supporting its selection for post-hoc analyses.Source: §4.2.1 · pp. 17–18
  • SSR = 1 and ACM > 0.5 characterized img_26, img_28, and Cosquer (Female) and El Castillo 25 (Male) across architectures and aggregation rules.Predictions agreed with prior hypotheses for img_26, Cosquer, and El Castillo 25, whereas confident predictions for img_28 and img_38 diverged from earlier interpretations. img_31 and Pech Merle produced conflicting classifications; lower confidence generally corresponded to historically ambiguous cases.Source: §4.2.1 · p. 18, Table 4 · p. 19

4.2.2. Post-hoc latent-space analysis and manifold mapping

  • Nine of nine stencils received matching sex labels from 2D-manifold k-NN classification and the EfficientNet-B3 ensemble.UMAP projections supported complementary neighborhood-based classification; predictions remained invariant across k ∈[1, 50), with SSR stabilizing for k ≥20 and final confidence averaged over [20, 50). img_26, img_28, Cosquer, and El Castillo 25 formed strongly supported neighborhoods, while the other five cases lay closer to inter-class overlap.Source: §4.2 · p. 17, §4.2.2 · p. 20, Table 5 · p. 21

4.2.3. Post-hoc interpretability analysis via spatial attribution

  • Across all nine stencils, LayerCAM relevance concentrated on fingers and interdigital spaces, with minimal central-palm or wrist activation.Each explanation geometrically aggregated 120 maps across silhouette variants and EfficientNet-B3 instances, suppressing contour artifacts and model stochasticity. The spatial pattern suggests reliance on finger geometry and relative spacing rather than global hand size or background structure.Source: Figure 11 · p. 21, §4.2.3 · p. 21
  • Compact attribution patterns accompanied strongly supported classifications, whereas diffuse patterns in Chauvet and Pech Merle aligned with interpretive ambiguity.Female-classified img_26, img_28, and Cosquer shared organized phalangeal and interdigital relevance, unlike Chauvet’s diffuse pattern and weaker latent-space support. Male cases generally emphasized metacarpophalangeal joints: img_38 and El Castillo 25 had well-defined clusters, while img_31 and Pech Merle showed broader activation.Source: §4.2.3 · pp. 21, 23, Table 6 · p. 22

5. Discussion

The framework propagates uncertainty and uses agreement across predictions, latent-space validation, and attribution analyses to assess classification stability. Its outputs remain probabilistic hypotheses conditioned on contemporary reference populations, not definitive determinations of prehistoric biological sex.

  • Uncertainty is deliberately introduced and propagated across processing, contours, silhouettes, representations, models, and decisions before hierarchical aggregation.Mechanisms include dual image processing, dual contour extraction, structured silhouette perturbations, multiple three-channel compositions, architectural diversity, stochastic replication, and ensemble fusion with silhouette support rates. Final predictions reflect convergence among controlled, imperfect observations rather than suppression of variability.Source: §5 · p. 23
  • 120 attribution instances per LayerCAM map support stable relevance patterns concentrated in anatomically meaningful finger and metacarpal regions, with comparatively little palm or wrist relevance.Geometric-median aggregation stabilizes explanations, while unsupervised UMAP followed by non-parametric k-NN tests class separability after aggressive dimensionality reduction. Highlighted interdigital spaces, proximal and middle phalanges, and metacarpal heads support biological plausibility, but attribution maps cannot establish causality.Source: §5 · p. 23
  • SSR ≈1 cases show localized, anatomically organized activations; agreement across predictions, latent-space validation, and attribution stability supports morphological stability, while divergence highlights ambiguity.Cases with architectural disagreement or reduced silhouette support show more diffuse attribution patterns. These converging analyses diagnose internal stability despite the absence of archaeological ground truth.Source: §5 · pp. 23–24
  • Archaeological interpretation should jointly evaluate predictions, confidence, latent-space organization, and attribution patterns, reporting low-confidence cases as indeterminate rather than forcing binary assignments.Robust candidates combine ensemble consistency, high silhouette support, large classification margins, stable manifold positioning, and ensemble–k-NN agreement; opposite indicators signal ambiguity. Ambiguity may reflect intermediate morphology, insufficient visual information, or reference-population limitations rather than classification failure.Source: §5 · p. 24
  • Unknown morphological differences between contemporary and Upper Paleolithic populations mean even highly consistent predictions remain reference-dependent probabilistic inferences rather than definitive sex determinations.Supervised methods necessarily assume some temporal stability of sexual dimorphism because archaeological sex labels are unavailable; that assumption cannot be directly verified. Future target-domain approaches remain constrained by archaeological sample quantity, quality, and representativeness.Source: §5 · p. 24
  • Future development should expand plausible-shape representations, replace point estimates with predictive distributions, and scale variability across pipeline stages toward a more fully probabilistic framework.Additional independent image-processing and manual-contour realizations could propagate more input variability. These extensions aim to distinguish data-driven ambiguity from model uncertainty and improve assessment of prediction stability.Source: §5 · pp. 24–25

6. Conclusions · Ethics approval

The framework models, propagates, and aggregates uncertainty to distinguish stable sex predictions from ambiguous prehistoric hand stencils. Convergent classification, latent-space, and interpretability evidence helps assess archaeological inference reliability without ground truth.

6. Conclusions

  • The framework combines controlled variability with hierarchical aggregation, deriving classifications from complementary observations rather than a single deterministic prediction.Uncertainty is modeled, propagated, and aggregated across image acquisition, manual annotation, silhouette representation, model training, and decision making.Source: §6 · p. 25
  • Applied to prehistoric hand stencils, the framework produced stable predictions for several cases while identifying intrinsically ambiguous cases through reduced silhouette support and lower classification margins.These confidence indicators help mitigate overinterpretation of uncertain archaeological evidence.Source: §6 · p. 25
  • Contemporary hand-silhouette experiments showed that deep learning ensembles reliably captured sexually dimorphic morphology, particularly in age groups where those traits were more strongly expressed.This age dependence qualifies the reported contemporary performance.Source: §6 · p. 25
  • Strong agreement among ensemble classification, unsupervised 2D latent-space validation, and aggregated LayerCAM explanations supported classifications with consistent representations and anatomically meaningful visual cues.This triangulation combines predictive, structural, and interpretability perspectives.Source: §6 · p. 25
  • Divergence across analyses exposes epistemic uncertainty and insufficient evidence for definitive attribution, making uncertainty informative for assessing archaeological inference reliability without ground truth.Such disagreements warrant cautious interpretation rather than definitive attribution.Source: §6 · p. 25
  • Future work should adopt more explicitly probabilistic formulations to represent input variability, quantify predictive uncertainty, and help distinguish morphological ambiguity from model-related uncertainty.These distinctions remain a proposed development rather than an established capability.Source: §6 · p. 25

Related papers

Loading 2608.14539v1…