Institution profile

Wuyi University

Academic institutionasia · cn
Official website
Research library6linked papers
Opportunities0open roles
Selected work

Representative Papers

Exact Linear Attention

May 13, 2026

This work addresses the limitations of conventional linear attention mechanisms, which suffer from approximation-induced errors, gradient explosion, and attention dilution. To overcome these issues, the authors propose a linear-complexity attention mechanism that eliminates approximation error by introducing novel kernel functions—such as the Hadamard Exp kernel and the squared Euclidean distance kernel—that satisfy non-negativity, discriminability, and geometric interpretability, enabling exact kernel decomposition. Furthermore, the model incorporates a Hyper Link structure, a Memory Lobe module, and a Mixture-of-Experts routing bias mechanism to enhance memory capacity, semantic alignment, and training stability. The resulting approach achieves efficient and accurate linear attention computation while preserving model performance and effectively mitigating gradient degradation and attention dilution.

0 citationsRead paper

Facial beauty prediction fusing transfer learning and broad learning system

Mar 13, 2026

Facial aesthetic prediction faces significant challenges due to data scarcity, high appearance variability, and the subjective nature of beauty, often leading to model overfitting and poor generalization. To address these issues, this work proposes a novel approach that integrates transfer learning with Broad Learning System (BLS), introducing EfficientNet-based BLS (E-BLS) and its enhanced variant with residual connections (ER-BLS). By optimizing the feature propagation architecture, the proposed models achieve improved predictive performance while maintaining computational efficiency during training. Extensive experiments demonstrate that both E-BLS and ER-BLS significantly outperform existing convolutional neural network (CNN) and BLS baselines across multiple evaluation metrics, thereby validating their effectiveness and strong generalization capability in facial aesthetic assessment.

0 citationsRead paper

PMT Waveform Simulation and Reconstruction with Conditional Diffusion Network

Feb 05, 2026

This work addresses the challenge of resolving individual photoelectrons in photomultiplier tube waveforms when nanosecond-scale multi-photon overlaps occur. To this end, the authors propose a weakly supervised learning framework based on a bidirectional conditional diffusion model that jointly optimizes waveform simulation and photoelectron sequence reconstruction. The method requires only raw waveforms and coarse photoelectron estimates—without ground-truth labels—and achieves high-fidelity reconstruction through bidirectional iterative training. Within the range of 1 to 5 photoelectrons, the approach attains a normalized photoelectron counting resolution of 99% and achieves 80% of the temporal resolution of fully supervised methods, significantly enhancing waveform reconstruction performance for photomultiplier tubes.

0 citationsRead paper
Recent publications

Latest Papers

Exact Linear Attention

May 13, 2026

This work addresses the limitations of conventional linear attention mechanisms, which suffer from approximation-induced errors, gradient explosion, and attention dilution. To overcome these issues, the authors propose a linear-complexity attention mechanism that eliminates approximation error by introducing novel kernel functions—such as the Hadamard Exp kernel and the squared Euclidean distance kernel—that satisfy non-negativity, discriminability, and geometric interpretability, enabling exact kernel decomposition. Furthermore, the model incorporates a Hyper Link structure, a Memory Lobe module, and a Mixture-of-Experts routing bias mechanism to enhance memory capacity, semantic alignment, and training stability. The resulting approach achieves efficient and accurate linear attention computation while preserving model performance and effectively mitigating gradient degradation and attention dilution.

0 citationsRead paper

Facial beauty prediction fusing transfer learning and broad learning system

Mar 13, 2026

Facial aesthetic prediction faces significant challenges due to data scarcity, high appearance variability, and the subjective nature of beauty, often leading to model overfitting and poor generalization. To address these issues, this work proposes a novel approach that integrates transfer learning with Broad Learning System (BLS), introducing EfficientNet-based BLS (E-BLS) and its enhanced variant with residual connections (ER-BLS). By optimizing the feature propagation architecture, the proposed models achieve improved predictive performance while maintaining computational efficiency during training. Extensive experiments demonstrate that both E-BLS and ER-BLS significantly outperform existing convolutional neural network (CNN) and BLS baselines across multiple evaluation metrics, thereby validating their effectiveness and strong generalization capability in facial aesthetic assessment.

0 citationsRead paper

PMT Waveform Simulation and Reconstruction with Conditional Diffusion Network

Feb 05, 2026

This work addresses the challenge of resolving individual photoelectrons in photomultiplier tube waveforms when nanosecond-scale multi-photon overlaps occur. To this end, the authors propose a weakly supervised learning framework based on a bidirectional conditional diffusion model that jointly optimizes waveform simulation and photoelectron sequence reconstruction. The method requires only raw waveforms and coarse photoelectron estimates—without ground-truth labels—and achieves high-fidelity reconstruction through bidirectional iterative training. Within the range of 1 to 5 photoelectrons, the approach attains a normalized photoelectron counting resolution of 99% and achieves 80% of the temporal resolution of fully supervised methods, significantly enhancing waveform reconstruction performance for photomultiplier tubes.

0 citationsRead paper