Selected work

Representative Papers

Quantum Temporal Convolutional Neural Networks for Cross-Sectional Equity Return Prediction: A Comparative Benchmark Study

Dec 06, 2025

Financial time series exhibit high noise levels, frequent regime shifts, and pose challenges for classical models due to limited generalization. To address these, this paper proposes the Quantum Temporal Convolutional Neural Network (QTCNN), which integrates a classical multi-scale temporal encoder with a parameter-efficient, trainable quantum convolutional circuit. Leveraging quantum superposition and entanglement, QTCNN enhances feature representation while mitigating overfitting. We evaluate the model on cross-sectional stock return prediction using JPX Tokyo Stock Exchange data and construct long–short portfolios for empirical validation. Out-of-sample testing yields a Sharpe ratio of 0.538—72% higher than the best classical baseline—demonstrating substantial improvements in prediction stability and generalization. This work pioneers the application of lightweight, differentiable quantum convolution to high-frequency quantitative forecasting, establishing a novel paradigm and providing empirical validation for practical quantum–classical hybrid modeling in real-world financial settings.

0 citationsRead paper

Ideal Observer for Segmentation of Dead Leaves Images

Dec 05, 2025

This work addresses the Bayesian optimal inference problem for image segmentation under the dead leaves model. Methodologically, it formulates and derives a theoretical framework for pixel grouping based on a Bayesian ideal observer, rigorously computing the posterior probability of segmentation given pixel observations using a joint prior over object location, shape, color, and texture; it further proposes a computationally tractable approximation scheme. Key contributions include: (i) establishing, for the first time, a principled performance upper bound for image segmentation in dead leaves occlusion scenarios; (ii) providing a unified theoretical benchmark to evaluate both human visual segmentation and computational algorithms; and (iii) empirically validating the feasibility and boundary behavior of this ideal observer on finite pixel sets. These results furnish foundational theoretical support for understanding visual grouping mechanisms and designing robust segmentation algorithms.

0 citationsRead paper

Revisiting Functional Derivatives in Multi-object Tracking

Aug 18, 2025

This work addresses the lack of mathematical rigor in defining functional derivatives of the Probability Generating Functional (PGFL) in multi-object tracking—particularly the reliance on heuristic arguments or ill-defined Dirac delta function operations. We propose a rigorous definition grounded in distribution theory and functional analysis, circumventing logical inconsistencies inherent in conventional formal derivations. Our approach systematically establishes fundamental properties of the PGFL functional derivative, including existence, uniqueness, and the chain rule, while clarifying equivalence relations and domain-specific validity among alternative definitions. The resulting theoretical framework significantly enhances the mathematical soundness and scalability of PGFL-based filters—such as the PHD and CBMeMBer filters—and provides a solid foundation for Bayesian multi-object filtering theory.

0 citationsRead paper

An Enhanced Focal Loss Function to Mitigate Class Imbalance in Auto Insurance Fraud Detection with Explainable AI

Aug 04, 2025

To address severe class imbalance in automobile insurance fraud detection, this paper proposes a dynamic multi-stage focal loss function. The method introduces a novel convex–nonconvex progressive switching mechanism that adaptively reweights hard samples during training, thereby mitigating local optima and slow convergence. Additionally, it integrates eXplainable AI (XAI) techniques to enhance model interpretability and facilitate actionable insights into fraud patterns. Experimental evaluation on a real-world automobile insurance dataset demonstrates that the proposed approach consistently outperforms standard focal loss and multiple baseline methods: accuracy, precision, recall, F1-score, and AUC all show significant improvement. These results validate the method’s robustness, effectiveness, and practical utility for imbalanced fraud detection tasks.

0 citationsRead paper

An approach to measuring the performance of Automatic Speech Recognition (ASR) models in the context of Large Language Model (LLM) powered applications

Jul 22, 2025

Conventional automatic speech recognition (ASR) evaluation using word error rate (WER) fails to capture the practical impact of ASR errors on downstream large language model (LLM)-driven tasks. Method: We propose a task-oriented ASR evaluation framework that (1) systematically classifies ASR error types and analyzes their contextual reparability within LLM prompts; (2) defines a multidimensional metric integrating semantic severity of errors, LLM-based correction success rate, and end-task completion accuracy; and (3) validates the framework empirically on representative speech-to-LLM pipelines—including voice command execution and meeting summary generation. Results: Our framework significantly outperforms WER in reflecting ASR effectiveness in real-world LLM applications. It provides an interpretable, quantifiable assessment grounded in downstream task performance, enabling principled, task-aware ASR model development and optimization.

0 citationsRead paper

Recent publications

Latest Papers

Quantum Temporal Convolutional Neural Networks for Cross-Sectional Equity Return Prediction: A Comparative Benchmark Study

Dec 06, 2025

Financial time series exhibit high noise levels, frequent regime shifts, and pose challenges for classical models due to limited generalization. To address these, this paper proposes the Quantum Temporal Convolutional Neural Network (QTCNN), which integrates a classical multi-scale temporal encoder with a parameter-efficient, trainable quantum convolutional circuit. Leveraging quantum superposition and entanglement, QTCNN enhances feature representation while mitigating overfitting. We evaluate the model on cross-sectional stock return prediction using JPX Tokyo Stock Exchange data and construct long–short portfolios for empirical validation. Out-of-sample testing yields a Sharpe ratio of 0.538—72% higher than the best classical baseline—demonstrating substantial improvements in prediction stability and generalization. This work pioneers the application of lightweight, differentiable quantum convolution to high-frequency quantitative forecasting, establishing a novel paradigm and providing empirical validation for practical quantum–classical hybrid modeling in real-world financial settings.

0 citationsRead paper

Ideal Observer for Segmentation of Dead Leaves Images

Dec 05, 2025

This work addresses the Bayesian optimal inference problem for image segmentation under the dead leaves model. Methodologically, it formulates and derives a theoretical framework for pixel grouping based on a Bayesian ideal observer, rigorously computing the posterior probability of segmentation given pixel observations using a joint prior over object location, shape, color, and texture; it further proposes a computationally tractable approximation scheme. Key contributions include: (i) establishing, for the first time, a principled performance upper bound for image segmentation in dead leaves occlusion scenarios; (ii) providing a unified theoretical benchmark to evaluate both human visual segmentation and computational algorithms; and (iii) empirically validating the feasibility and boundary behavior of this ideal observer on finite pixel sets. These results furnish foundational theoretical support for understanding visual grouping mechanisms and designing robust segmentation algorithms.

0 citationsRead paper

Revisiting Functional Derivatives in Multi-object Tracking

Aug 18, 2025

This work addresses the lack of mathematical rigor in defining functional derivatives of the Probability Generating Functional (PGFL) in multi-object tracking—particularly the reliance on heuristic arguments or ill-defined Dirac delta function operations. We propose a rigorous definition grounded in distribution theory and functional analysis, circumventing logical inconsistencies inherent in conventional formal derivations. Our approach systematically establishes fundamental properties of the PGFL functional derivative, including existence, uniqueness, and the chain rule, while clarifying equivalence relations and domain-specific validity among alternative definitions. The resulting theoretical framework significantly enhances the mathematical soundness and scalability of PGFL-based filters—such as the PHD and CBMeMBer filters—and provides a solid foundation for Bayesian multi-object filtering theory.

0 citationsRead paper

An Enhanced Focal Loss Function to Mitigate Class Imbalance in Auto Insurance Fraud Detection with Explainable AI

Aug 04, 2025

To address severe class imbalance in automobile insurance fraud detection, this paper proposes a dynamic multi-stage focal loss function. The method introduces a novel convex–nonconvex progressive switching mechanism that adaptively reweights hard samples during training, thereby mitigating local optima and slow convergence. Additionally, it integrates eXplainable AI (XAI) techniques to enhance model interpretability and facilitate actionable insights into fraud patterns. Experimental evaluation on a real-world automobile insurance dataset demonstrates that the proposed approach consistently outperforms standard focal loss and multiple baseline methods: accuracy, precision, recall, F1-score, and AUC all show significant improvement. These results validate the method’s robustness, effectiveness, and practical utility for imbalanced fraud detection tasks.

0 citationsRead paper

An approach to measuring the performance of Automatic Speech Recognition (ASR) models in the context of Large Language Model (LLM) powered applications

Jul 22, 2025

Conventional automatic speech recognition (ASR) evaluation using word error rate (WER) fails to capture the practical impact of ASR errors on downstream large language model (LLM)-driven tasks. Method: We propose a task-oriented ASR evaluation framework that (1) systematically classifies ASR error types and analyzes their contextual reparability within LLM prompts; (2) defines a multidimensional metric integrating semantic severity of errors, LLM-based correction success rate, and end-task completion accuracy; and (3) validates the framework empirically on representative speech-to-LLM pipelines—including voice command execution and meeting summary generation. Results: Our framework significantly outperforms WER in reflecting ASR effectiveness in real-world LLM applications. It provides an interpretable, quantifiable assessment grounded in downstream task performance, enabling principled, task-aware ASR model development and optimization.

0 citationsRead paper