Selected work

Representative Papers

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

Aug 14, 2026

This study addresses the uncertainty arising from missing ground truth and image degradation in sex determination of prehistoric hand stencils by proposing an end-to-end uncertainty-aware deep learning framework. By integrating ensemble learning, contour enhancement, and manifold mapping, the method explicitly models and propagates uncertainty throughout the analytical pipeline while employing explainable AI to verify anatomical consistency, thereby transforming uncertainty into a quantifiable component of archaeological inference. Achieving over 88% accuracy on modern samples, the framework effectively distinguishes between morphologically stable and ambiguous cases. Consequently, this work enables robust decoding, confidence quantification, and reproducible inference regarding sex attribution in prehistoric rock art, offering a rigorous computational approach to mitigating epistemic limitations in paleoanthropological research.

0 citationsRead paper

Emergence of Statistical Financial Factors by a Diffusion Process

Apr 13, 2026

This study addresses the limitation of traditional factor models, which rely on exogenously specified factors and fail to capture endogenous interaction mechanisms among assets. The authors propose a network-coupled map-based model of financial markets that generates statistical factors endogenously through dynamic asset interactions. By applying an orthogonal transformation of the graph Laplacian matrix and employing center manifold dimensionality reduction, the model reveals an intrinsic link between initial asset clustering structures and the emergent number of factors. Furthermore, it integrates coupled iterative maps with network diffusion dynamics to simulate the impact of irrational trading behavior on asset prices. Empirical experiments demonstrate that, within an optimal parameter regime, the endogenously generated factors effectively explain cross-sectional variance in asset returns, thereby validating both the feasibility and explanatory power of the proposed interaction-driven factor mechanism.

0 citationsRead paper

Sign language recognition from skeletal data using graph and recurrent neural networks

Nov 08, 2025

This work addresses isolated sign language gesture recognition by proposing a graph-temporal joint modeling approach based on skeletal pose sequences. To simultaneously capture the spatial topological structure among joints and the dynamic temporal evolution of gestures, we design Graph-GRU: a hybrid architecture integrating graph neural networks (GNNs) to model spatial dependencies in the skeleton topology and gated recurrent units (GRUs) to encode long-range temporal dynamics. The framework enables end-to-end learning of pose-driven spatiotemporal feature representations and exhibits strong scalability. Extensive experiments on the large-scale AUTSL sign language dataset demonstrate significant improvements in classification accuracy over state-of-the-art methods. Results validate the effectiveness of jointly leveraging structural priors (via graph-based spatial modeling) and sequential dynamics (via recurrent temporal modeling) for enhancing sign language recognition performance. This work establishes a novel paradigm for pose-based sign language understanding.

0 citationsRead paper

Recent publications

Latest Papers

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

Aug 14, 2026

This study addresses the uncertainty arising from missing ground truth and image degradation in sex determination of prehistoric hand stencils by proposing an end-to-end uncertainty-aware deep learning framework. By integrating ensemble learning, contour enhancement, and manifold mapping, the method explicitly models and propagates uncertainty throughout the analytical pipeline while employing explainable AI to verify anatomical consistency, thereby transforming uncertainty into a quantifiable component of archaeological inference. Achieving over 88% accuracy on modern samples, the framework effectively distinguishes between morphologically stable and ambiguous cases. Consequently, this work enables robust decoding, confidence quantification, and reproducible inference regarding sex attribution in prehistoric rock art, offering a rigorous computational approach to mitigating epistemic limitations in paleoanthropological research.

0 citationsRead paper

Emergence of Statistical Financial Factors by a Diffusion Process

Apr 13, 2026

This study addresses the limitation of traditional factor models, which rely on exogenously specified factors and fail to capture endogenous interaction mechanisms among assets. The authors propose a network-coupled map-based model of financial markets that generates statistical factors endogenously through dynamic asset interactions. By applying an orthogonal transformation of the graph Laplacian matrix and employing center manifold dimensionality reduction, the model reveals an intrinsic link between initial asset clustering structures and the emergent number of factors. Furthermore, it integrates coupled iterative maps with network diffusion dynamics to simulate the impact of irrational trading behavior on asset prices. Empirical experiments demonstrate that, within an optimal parameter regime, the endogenously generated factors effectively explain cross-sectional variance in asset returns, thereby validating both the feasibility and explanatory power of the proposed interaction-driven factor mechanism.

0 citationsRead paper

Sign language recognition from skeletal data using graph and recurrent neural networks

Nov 08, 2025

This work addresses isolated sign language gesture recognition by proposing a graph-temporal joint modeling approach based on skeletal pose sequences. To simultaneously capture the spatial topological structure among joints and the dynamic temporal evolution of gestures, we design Graph-GRU: a hybrid architecture integrating graph neural networks (GNNs) to model spatial dependencies in the skeleton topology and gated recurrent units (GRUs) to encode long-range temporal dynamics. The framework enables end-to-end learning of pose-driven spatiotemporal feature representations and exhibits strong scalability. Extensive experiments on the large-scale AUTSL sign language dataset demonstrate significant improvements in classification accuracy over state-of-the-art methods. Results validate the effectiveness of jointly leveraging structural priors (via graph-based spatial modeling) and sequential dynamics (via recurrent temporal modeling) for enhancing sign language recognition performance. This work establishes a novel paradigm for pose-based sign language understanding.

0 citationsRead paper