摘要
针对RMSprop算法存在梯度消失和局部最优的问题,本文提出了一种基于Borges差分的RMSprop算法并应用到卷积神经网络参数训练方法.根据Borges分形导数的定义,本文给出了Borges差分的定义;将Borges差分与RMSprop算法相结合,提出了基于Borges差分的RMSprop优化算法,提高了图像识别的精度和学习收敛速度;给出了一种基于梯度信息的自适应的调整阶次的方法.本文以Fashion-MNIST数据集为例,分析了阶次对网络参数训练结果的影响,验证了本文提出的算法提高卷积神经网络的识别精度和学习收敛速度的有效性.
In order to solve the problem of gradient vanishing and local optimum in RMSprop algorithm,the RMSprop algorithm based on Borges difference is proposed to train parameters of convolutional neural networks.According to the definition of Borges fractal derivative,the definition of Borges difference is provided.Combining Borges difference with RMSprop algorithm,an optimization algorithm of RMSprop based on Borges difference is proposed,which improves the accuracy and learning convergence speed of image recognition.This paper presents an adaptive method to adjust the order based on gradient information.Taking the Fashion-MNIST data set as an example,the influence of the order on the training results of the network parameters is analyzed,and the effectiveness of the proposed algorithm to improve the recognition accuracy and learning convergence speed of the convolutional neural network is verified.
作者
高哲
剪静
GAO Zhe;JIAN Jing(College of Light Industry,Liaoning University,Shenyang 110036,China;School of Mathematics and Statistics,Liaoning University,Shenyang 110036,China)
出处
《辽宁大学学报（自然科学版）》
CAS
2023年第1期1-9,F0002,共10页
Journal of Liaoning University：Natural Sciences Edition
基金
辽宁省教育厅基础研究项目(LJC202010)
辽宁省自然科学基金项目(20180520009)
兴辽英才计划项目(XLYC1807229)
沈阳市中青年科技创新人才支持计划项目(RC210082)。
