Bypass Observation: A Conceptual Design of a Non-Intrusive Layer-Wise Semantic Extraction Architecture
为了解决大语言模型内部推理过程不可见的问题，本文提出了一种非侵入式的逐层语义提取架构——旁路观察，通过附加只读观察头到选定的Transformer层来实现。
0 citationsRead paper
Selected work
Representative Papers
Medical Image Segmentation based on Deep Active Contour and Mean Curvature Loss Function
This work addresses the common issue in medical image segmentation where insufficient geometric priors lead to poor boundary representation. To this end, it introduces mean curvature as a geometric constraint into deep active contour models for the first time, formulating a novel loss function to enhance geometric consistency of segmentation boundaries. A lightweight convolutional kernel is employed to approximate mean curvature efficiently, significantly reducing computational overhead. Furthermore, the method integrates the Chan–Vese model with convolutional neural networks to jointly optimize region and boundary information. Evaluated on liver and spleen datasets, the proposed approach achieves state-of-the-art performance, markedly improving both segmentation accuracy and geometric plausibility.
0 citationsRead paper
Gradient Perturbation: Learning to Perturb Gradients for Adaptive Training
This work addresses the limited exploration of gradient perturbations during backpropagation in existing training methods, which often focus solely on forward perturbations and lack mechanisms for class-aware adaptive optimization. To bridge this gap, we establish a unified theoretical framework for gradient perturbation that subsumes techniques such as Sharpness-Aware Minimization (SAM) and gradient clipping. Within this framework, we propose Label-aware Perturbed Gradients (LPG), a novel method that adaptively modulates gradient norms to induce class-conditional augmentation effects. Leveraging PAC-Bayesian analysis, we theoretically connect the magnitude of gradient perturbations to generalization performance. LPG is designed as a modular, plug-and-play component and demonstrates consistent superiority over existing approaches across balanced classification, long-tailed learning, and noisy-label settings, while seamlessly integrating with other training strategies.
0 citationsRead paper
Learning to Perturb Hidden Representations for Generalizable Deep Learning
Existing deep learning approaches lack a unified and effective mechanism for perturbing hidden-layer activations, limiting their ability to systematically enhance model generalization. This work proposes a class-aware, learnable perturbation framework for activations (LPA), which, for the first time, systematically analyzes the underlying mechanisms of hidden activation perturbation and reveals the opposing effects of expansive versus contractive perturbations on generalization. By leveraging projected gradient descent (PGD) optimization, LPA enables class-adaptive perturbations and theoretically connects them to flat minima and inter-layer amplification effects. Extensive experiments demonstrate that LPA significantly outperforms current methods across balanced classification, long-tailed learning, and domain generalization tasks, and further exhibits complementary gains when combined with logit-level perturbation techniques such as LPL.
0 citationsRead paper
StomaD2: An All-in-One System for Intelligent Stomatal Phenotype Analysis via Diffusion-Based Restoration Detection Network
Traditional stomatal phenotyping relies on destructive sampling and manual annotation, which are ill-suited for high-throughput, non-invasive, and large-scale field applications. This work proposes an intelligent analysis system that integrates diffusion-based image restoration with rotation-aware object detection. The system employs a diffusion model to recover degraded images and introduces a specialized detection network tailored for small, densely packed stomata in complex backgrounds. Key innovations include column-wise global feature interaction, context-aware resampling and reweighting mechanisms, and a feature rearrangement module, collectively enhancing detection accuracy and robustness. Evaluated on maize and wheat datasets, the system achieves accuracies of 0.994 and 0.992, respectively, with an F1-score/mAP of 0.989. It enables rapid extraction of eight stomatal phenotypic traits and demonstrates strong generalization across more than 130 plant species.
0 citationsRead paper
Recent publications
Latest Papers
Bypass Observation: A Conceptual Design of a Non-Intrusive Layer-Wise Semantic Extraction Architecture
为了解决大语言模型内部推理过程不可见的问题，本文提出了一种非侵入式的逐层语义提取架构——旁路观察，通过附加只读观察头到选定的Transformer层来实现。
0 citationsRead paper
Medical Image Segmentation based on Deep Active Contour and Mean Curvature Loss Function
This work addresses the common issue in medical image segmentation where insufficient geometric priors lead to poor boundary representation. To this end, it introduces mean curvature as a geometric constraint into deep active contour models for the first time, formulating a novel loss function to enhance geometric consistency of segmentation boundaries. A lightweight convolutional kernel is employed to approximate mean curvature efficiently, significantly reducing computational overhead. Furthermore, the method integrates the Chan–Vese model with convolutional neural networks to jointly optimize region and boundary information. Evaluated on liver and spleen datasets, the proposed approach achieves state-of-the-art performance, markedly improving both segmentation accuracy and geometric plausibility.
0 citationsRead paper
Gradient Perturbation: Learning to Perturb Gradients for Adaptive Training
This work addresses the limited exploration of gradient perturbations during backpropagation in existing training methods, which often focus solely on forward perturbations and lack mechanisms for class-aware adaptive optimization. To bridge this gap, we establish a unified theoretical framework for gradient perturbation that subsumes techniques such as Sharpness-Aware Minimization (SAM) and gradient clipping. Within this framework, we propose Label-aware Perturbed Gradients (LPG), a novel method that adaptively modulates gradient norms to induce class-conditional augmentation effects. Leveraging PAC-Bayesian analysis, we theoretically connect the magnitude of gradient perturbations to generalization performance. LPG is designed as a modular, plug-and-play component and demonstrates consistent superiority over existing approaches across balanced classification, long-tailed learning, and noisy-label settings, while seamlessly integrating with other training strategies.
0 citationsRead paper
Learning to Perturb Hidden Representations for Generalizable Deep Learning
Existing deep learning approaches lack a unified and effective mechanism for perturbing hidden-layer activations, limiting their ability to systematically enhance model generalization. This work proposes a class-aware, learnable perturbation framework for activations (LPA), which, for the first time, systematically analyzes the underlying mechanisms of hidden activation perturbation and reveals the opposing effects of expansive versus contractive perturbations on generalization. By leveraging projected gradient descent (PGD) optimization, LPA enables class-adaptive perturbations and theoretically connects them to flat minima and inter-layer amplification effects. Extensive experiments demonstrate that LPA significantly outperforms current methods across balanced classification, long-tailed learning, and domain generalization tasks, and further exhibits complementary gains when combined with logit-level perturbation techniques such as LPL.
0 citationsRead paper
StomaD2: An All-in-One System for Intelligent Stomatal Phenotype Analysis via Diffusion-Based Restoration Detection Network
Traditional stomatal phenotyping relies on destructive sampling and manual annotation, which are ill-suited for high-throughput, non-invasive, and large-scale field applications. This work proposes an intelligent analysis system that integrates diffusion-based image restoration with rotation-aware object detection. The system employs a diffusion model to recover degraded images and introduces a specialized detection network tailored for small, densely packed stomata in complex backgrounds. Key innovations include column-wise global feature interaction, context-aware resampling and reweighting mechanisms, and a feature rearrangement module, collectively enhancing detection accuracy and robustness. Evaluated on maize and wheat datasets, the system achieves accuracies of 0.994 and 0.992, respectively, with an F1-score/mAP of 0.989. It enables rapid extraction of eight stomatal phenotypic traits and demonstrates strong generalization across more than 130 plant species.
0 citationsRead paper