EFQ-Softmax: Exp-Free Quantization for Softmax
本文提出EFQ-Softmax方法，直接生成低比特概率，解决softmax路径中高精度概率生产者与低比特矩阵消费者之间的不匹配问题。
0 citationsRead paper
Selected work
Representative Papers
Effective Attention-Guided Multi-Scale Medical Network for Skin Lesion Segmentation
To address the challenges of irregular lesion morphology and low contrast in skin lesion segmentation, this paper proposes a novel encoder-decoder network. Methodologically, it introduces a multi-scale residual encoder with a multi-resolution, multi-channel fusion (MRCF) module to enhance cross-scale feature representation; designs a cross-mixed attention module (CMAM) to dynamically recalibrate attention scope; and constructs an external attention bridge (EAB) to mitigate information loss in U-Net skip connections. Evaluated on multiple public skin lesion segmentation benchmarks, the method consistently outperforms state-of-the-art CNN- and Transformer-based models, achieving average Dice score improvements of 2.3–4.1%. Notably, it demonstrates superior robustness on small lesions and low-contrast regions. The core contribution lies in the synergistic optimization of feature granularity, attention mechanisms, and information flow pathways, establishing a new, interpretable, and high-accuracy paradigm for medical image segmentation.
0 citationsRead paper
$L_1$-norm Regularized Indefinite Kernel Logistic Regression
To address the challenge of simultaneously achieving sparsity, interpretability, and generalization in indefinite kernel logistic regression (IKLR), this paper introduces the $L_1$-norm regularization into the IKLR framework for the first time, yielding the Sparse Indefinite Kernel Logistic Regression (S-IKLR) model. To tackle the resulting nonsmooth and nonconvex optimization problem, we propose an efficient proximal linearization-based algorithm with theoretical convergence guarantees. S-IKLR leverages the expressive power of indefinite kernels to capture complex data structures while enforcing sparsity via $L_1$ regularization, thereby substantially reducing the number of nonzero parameters. Extensive experiments on multiple benchmark datasets demonstrate that S-IKLR achieves superior classification accuracy compared to state-of-the-art IKLR and sparse kernel methods. Moreover, it attains 30–60% higher model sparsity, leading to significantly improved interpretability and generalization performance.
0 citationsRead paper
Recent publications
Latest Papers
EFQ-Softmax: Exp-Free Quantization for Softmax
本文提出EFQ-Softmax方法，直接生成低比特概率，解决softmax路径中高精度概率生产者与低比特矩阵消费者之间的不匹配问题。
0 citationsRead paper
Effective Attention-Guided Multi-Scale Medical Network for Skin Lesion Segmentation
To address the challenges of irregular lesion morphology and low contrast in skin lesion segmentation, this paper proposes a novel encoder-decoder network. Methodologically, it introduces a multi-scale residual encoder with a multi-resolution, multi-channel fusion (MRCF) module to enhance cross-scale feature representation; designs a cross-mixed attention module (CMAM) to dynamically recalibrate attention scope; and constructs an external attention bridge (EAB) to mitigate information loss in U-Net skip connections. Evaluated on multiple public skin lesion segmentation benchmarks, the method consistently outperforms state-of-the-art CNN- and Transformer-based models, achieving average Dice score improvements of 2.3–4.1%. Notably, it demonstrates superior robustness on small lesions and low-contrast regions. The core contribution lies in the synergistic optimization of feature granularity, attention mechanisms, and information flow pathways, establishing a new, interpretable, and high-accuracy paradigm for medical image segmentation.
0 citationsRead paper
$L_1$-norm Regularized Indefinite Kernel Logistic Regression
To address the challenge of simultaneously achieving sparsity, interpretability, and generalization in indefinite kernel logistic regression (IKLR), this paper introduces the $L_1$-norm regularization into the IKLR framework for the first time, yielding the Sparse Indefinite Kernel Logistic Regression (S-IKLR) model. To tackle the resulting nonsmooth and nonconvex optimization problem, we propose an efficient proximal linearization-based algorithm with theoretical convergence guarantees. S-IKLR leverages the expressive power of indefinite kernels to capture complex data structures while enforcing sparsity via $L_1$ regularization, thereby substantially reducing the number of nonzero parameters. Extensive experiments on multiple benchmark datasets demonstrate that S-IKLR achieves superior classification accuracy compared to state-of-the-art IKLR and sparse kernel methods. Moreover, it attains 30–60% higher model sparsity, leading to significantly improved interpretability and generalization performance.
0 citationsRead paper