期刊
NEUROCOMPUTING
卷 477, 期 -, 页码 36-45出版社
ELSEVIER
DOI: 10.1016/j.neucom.2022.01.006
关键词
Caputo fractional-order gradient method; Population extremal optimization; Initial bias and weight; MNIST dataset; Fractional-order convolutional neural networks
资金
- National Natural Science Foundation of China [61973102, 61972288, 61871427]
- Zhejiang Provincial Natural Science Foundation of China [LZ22F010003]
This article focuses on the intelligent optimization issue using PEO-FOCNN, which combines fractional order convolutional neural networks (FOCNNs) with population extremal optimization (PEO). The Caputo fractional-order gradient method (CFOGM) is introduced to improve the optimization performance of FOCNN. The experiments demonstrate the superiority of PEO-FOCNN over other optimization algorithms on the MNIST dataset.
This article is devoted to the intelligent optimization issue by means of PEO-FOCNN, i.e., the fractional order convolutional neural networks (FOCNNs) with population extremal optimization (PEO). The Caputo fractional-order gradient method (CFOGM) is adopted to improve the dynamic updating effectiveness of the biases and weights for convolutional neural networks (CNN). Moreover, considering the significance of the initial biases and weights and their updating mechanisms to the optimization performance of FOCNN, the PEO algorithm is used to seek an optimal selection from lots of the initial biases and weights. The optimization effect of PEO method for FOCNN is demonstrated by the training and testing accuracies of PEO-FOCNN compared with standard FOCNN. And, the superiority of the proposed PEO-FOCNN to FOCNN based on some other popular optimization algorithms, such as the genetic algorithm-based FOCNN (GA-FOCNN), differential evolution-based FOCNN (DE-FOCNN) and particle swarm optimization-based FOCNN (PSO-FOCNN), is verified by the experiments on the MNIST dataset in terms of three types of statistical tests. (c) 2022 Elsevier B.V. All rights reserved.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据