Selected work

Representative Papers

LabelFusion-TS: Fusing Large Language Models, Transformer Encoders, and Financial Time Series for Monetary-Policy Stance Classification

Aug 12, 2026

This work addresses the limitations of existing financial text classification methods, which often neglect market context and struggle to accurately discern hawkish, dovish, or neutral stances in Federal Reserve communications. To overcome this, the authors propose LabelFusion-TS, a novel system that, for the first time, incorporates financial market time series as an auxiliary modality. The approach fuses a fine-tuned RoBERTa model, prompt-driven large language models, and a time series Transformer, employing a two-stage training strategy to mitigate the scarcity of labeled data. Evaluated on a test set spanning 2015–2022, the model achieves a weighted F1 score of 70.2% using only 240 manually annotated samples—significantly outperforming zero-shot large language models (64.1%)—thereby demonstrating the efficacy of multimodal fusion and few-shot learning in this domain.

0 citationsRead paper

No Attention, No Problem: DPU-Aware Attention Approximation in Modern YOLO on FPGA

Jul 14, 2026

This work addresses the challenges of deploying attention-enhanced YOLO models on edge FPGAs, where limited DPU support for certain operators—such as specific activation functions, split operations, and spatial attention mechanisms—hinders simultaneous optimization of accuracy, speed, and energy efficiency. Focusing on AMD FPGA platforms, the study proposes a DPU-aware architectural optimization strategy for YOLO, incorporating operator substitution, replacement of split operations with 1×1 convolutions, DPU-compatible approximations of attention mechanisms, INT8 quantization, and multi-configuration deployment. This approach enables, for the first time, efficient end-to-end inference of modern YOLO variants—including YOLOv26 and YOLOv11—on the Xilinx ZCU104. Experiments demonstrate that YOLOv26n achieves 34.05 and 29.55 FPS on six benchmark datasets for standard and oriented object detection, respectively, with approximately threefold lower power consumption and only a ~5% mAP degradation.

0 citationsRead paper

Estimation of Motor Unit Parameters from Surface Electromyograms using an Informed Autoencoder

May 08, 2026

This study addresses the challenges of non-invasive, simultaneous estimation of personalized motor unit parameters—such as innervation zone center and conduction velocity—which are hindered by modeling complexity and reliance on manual intervention. The authors propose a physics-informed autoencoder, applied for the first time to surface electromyography (sEMG) signal analysis, that jointly reconstructs the observed signals and inversely infers multiple biophysical parameters within its latent space. By synergistically integrating data-driven learning with mechanistic modeling, the method substantially reduces the manual effort typically required in white-box modeling approaches. Evaluated on synthetic data, the approach achieves high accuracy, yielding a mean error of 2.60 mm for innervation zone center localization and 0.17 m/s for conduction velocity estimation, thereby demonstrating both precision and feasibility.

0 citationsRead paper

Residual-Informed Learning of Solutions to Algebraic Loops

Oct 10, 2025

To address numerical ambiguity and inefficiency arising from algebraic loops in equation-based modeling languages (e.g., Modelica), this paper proposes an unsupervised, residual-driven neural surrogate modeling method. A feedforward neural network serves as the surrogate, with the algebraic loop residual explicitly incorporated into the loss function—enabling end-to-end, label-free training that eliminates reliance on supervised datasets. Crucially, residual minimization resolves solution multiplicity by steering convergence toward physically consistent solutions—not statistical averages. An integrated error-control strategy ensures numerical accuracy. Evaluated on the IEEE 14-Bus system, the method achieves a 60% speedup over conventional numerical solvers while maintaining comparable precision. The core innovation lies in the first use of algebraic loop residuals as an unsupervised learning signal for neural networks, uniquely balancing robustness, interpretability, and engineering practicality.

0 citationsRead paper

Efficient Training of Physics-enhanced Neural ODEs via Direct Collocation and Nonlinear Programming

May 06, 2025

To address the poor stability, high computational cost, and low accuracy of ODE solvers in training physics-enhanced neural ordinary differential equations (PeNODEs), this work reformulates training as a dynamic optimization problem. It employs a high-order implicit Runge–Kutta method—based on flipped Legendre–Gauss–Radau collocation points—for temporal discretization, yielding a large-scale nonlinear programming (NLP) problem solved efficiently via Ipopt. Our approach achieves, for the first time, joint optimization of neural parameters and state trajectories; extends direct collocation to incorporate physics-based constraints; and releases an open-source implementation supporting parallel computation and interoperability with OpenModelica, while also accommodating Neural DAE training. Experiments on the Quarter Vehicle and Van der Pol systems demonstrate substantial improvements in accuracy and training speed, superior generalization with smaller networks, and enhanced numerical robustness—particularly for stiff systems.

0 citationsRead paper

Recent publications

Latest Papers

LabelFusion-TS: Fusing Large Language Models, Transformer Encoders, and Financial Time Series for Monetary-Policy Stance Classification

Aug 12, 2026

This work addresses the limitations of existing financial text classification methods, which often neglect market context and struggle to accurately discern hawkish, dovish, or neutral stances in Federal Reserve communications. To overcome this, the authors propose LabelFusion-TS, a novel system that, for the first time, incorporates financial market time series as an auxiliary modality. The approach fuses a fine-tuned RoBERTa model, prompt-driven large language models, and a time series Transformer, employing a two-stage training strategy to mitigate the scarcity of labeled data. Evaluated on a test set spanning 2015–2022, the model achieves a weighted F1 score of 70.2% using only 240 manually annotated samples—significantly outperforming zero-shot large language models (64.1%)—thereby demonstrating the efficacy of multimodal fusion and few-shot learning in this domain.

0 citationsRead paper

No Attention, No Problem: DPU-Aware Attention Approximation in Modern YOLO on FPGA

Jul 14, 2026

This work addresses the challenges of deploying attention-enhanced YOLO models on edge FPGAs, where limited DPU support for certain operators—such as specific activation functions, split operations, and spatial attention mechanisms—hinders simultaneous optimization of accuracy, speed, and energy efficiency. Focusing on AMD FPGA platforms, the study proposes a DPU-aware architectural optimization strategy for YOLO, incorporating operator substitution, replacement of split operations with 1×1 convolutions, DPU-compatible approximations of attention mechanisms, INT8 quantization, and multi-configuration deployment. This approach enables, for the first time, efficient end-to-end inference of modern YOLO variants—including YOLOv26 and YOLOv11—on the Xilinx ZCU104. Experiments demonstrate that YOLOv26n achieves 34.05 and 29.55 FPS on six benchmark datasets for standard and oriented object detection, respectively, with approximately threefold lower power consumption and only a ~5% mAP degradation.

0 citationsRead paper

Estimation of Motor Unit Parameters from Surface Electromyograms using an Informed Autoencoder

May 08, 2026

This study addresses the challenges of non-invasive, simultaneous estimation of personalized motor unit parameters—such as innervation zone center and conduction velocity—which are hindered by modeling complexity and reliance on manual intervention. The authors propose a physics-informed autoencoder, applied for the first time to surface electromyography (sEMG) signal analysis, that jointly reconstructs the observed signals and inversely infers multiple biophysical parameters within its latent space. By synergistically integrating data-driven learning with mechanistic modeling, the method substantially reduces the manual effort typically required in white-box modeling approaches. Evaluated on synthetic data, the approach achieves high accuracy, yielding a mean error of 2.60 mm for innervation zone center localization and 0.17 m/s for conduction velocity estimation, thereby demonstrating both precision and feasibility.

0 citationsRead paper

Residual-Informed Learning of Solutions to Algebraic Loops

Oct 10, 2025

To address numerical ambiguity and inefficiency arising from algebraic loops in equation-based modeling languages (e.g., Modelica), this paper proposes an unsupervised, residual-driven neural surrogate modeling method. A feedforward neural network serves as the surrogate, with the algebraic loop residual explicitly incorporated into the loss function—enabling end-to-end, label-free training that eliminates reliance on supervised datasets. Crucially, residual minimization resolves solution multiplicity by steering convergence toward physically consistent solutions—not statistical averages. An integrated error-control strategy ensures numerical accuracy. Evaluated on the IEEE 14-Bus system, the method achieves a 60% speedup over conventional numerical solvers while maintaining comparable precision. The core innovation lies in the first use of algebraic loop residuals as an unsupervised learning signal for neural networks, uniquely balancing robustness, interpretability, and engineering practicality.

0 citationsRead paper

Efficient Training of Physics-enhanced Neural ODEs via Direct Collocation and Nonlinear Programming

May 06, 2025

To address the poor stability, high computational cost, and low accuracy of ODE solvers in training physics-enhanced neural ordinary differential equations (PeNODEs), this work reformulates training as a dynamic optimization problem. It employs a high-order implicit Runge–Kutta method—based on flipped Legendre–Gauss–Radau collocation points—for temporal discretization, yielding a large-scale nonlinear programming (NLP) problem solved efficiently via Ipopt. Our approach achieves, for the first time, joint optimization of neural parameters and state trajectories; extends direct collocation to incorporate physics-based constraints; and releases an open-source implementation supporting parallel computation and interoperability with OpenModelica, while also accommodating Neural DAE training. Experiments on the Quarter Vehicle and Van der Pol systems demonstrate substantial improvements in accuracy and training speed, superior generalization with smaller networks, and enhanced numerical robustness—particularly for stiff systems.

0 citationsRead paper