Scholarlink

<p>Activation functions (AFs) are critical components in neural networks, as they provide non-linearity and allow the models to recognize complex patterns. However, traditional sigmoid AFs often face limitations, particularly in deep architectures, due to saturation at extreme values that hinders learning. In this paper, we propose a new N-degree Mittag-Leffler Sigmoidal Activation Function (N-MLSAF), which enhances the sigmoid function by generalizing its exponential behavior. The two adjustable parameters of the N-MLSAF allow for dynamic adaptation of the function’s shape, thereby facilitating a broader output range and smoother learning. We integrate the N-MLSAF into several network architectures, including LeNet-5 and ResNet-20, and compare its performance against widely used AFs such as Sigmoid, Tanh, Gish, Mish, E-tanh, E-swish, and Swish, across benchmark datasets such as MNIST, FMNIST, CIFAR-10 and CIFAR-100. The results demonstrate that the 1-MLSAF consistently outperforms these existing functions, achieving accuracies of up to 99.48%. This work addresses the limitations of standard AFs exploring new opportunities for the advancement of deep learning and fractional calculus. To access the codes of the proposed N-MLSAF: <a href="https://github.com/sivaranjani2024-cyber/N-MLSAF-Activation">https://github.com/sivaranjani2024-cyber/N-MLSAF-Activation</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A novel N-degree mittag-leffler sigmoidal activation function for deep learning algorithms

  • 期刊论文

  • 发表日期：

    2026年5月14日

查看全文
  • V. Sivaranjani,
  • N. Ramesh Babu,
  • N. Padmaja

摘要

Activation functions (AFs) are critical components in neural networks, as they provide non-linearity and allow the models to recognize complex patterns. However, traditional sigmoid AFs often face limitations, particularly in deep architectures, due to saturation at extreme values that hinders learning. In this paper, we propose a new N-degree Mittag-Leffler Sigmoidal Activation Function (N-MLSAF), which enhances the sigmoid function by generalizing its exponential behavior. The two adjustable parameters of the N-MLSAF allow for dynamic adaptation of the function’s shape, thereby facilitating a broader output range and smoother learning. We integrate the N-MLSAF into several network architectures, including LeNet-5 and ResNet-20, and compare its performance against widely used AFs such as Sigmoid, Tanh, Gish, Mish, E-tanh, E-swish, and Swish, across benchmark datasets such as MNIST, FMNIST, CIFAR-10 and CIFAR-100. The results demonstrate that the 1-MLSAF consistently outperforms these existing functions, achieving accuracies of up to 99.48%. This work addresses the limitations of standard AFs exploring new opportunities for the advancement of deep learning and fractional calculus. To access the codes of the proposed N-MLSAF: https://github.com/sivaranjani2024-cyber/N-MLSAF-Activation.