基于ELM-AE和BP算法的极限学习机特征表示方法  

Feature representation method for extreme learning machine based on ELM-AE and BP algorithms

作　　者：苗军 刘晓 常艺茹 乔元华[2] MIAO Jun;LIU Xiao;CHANG Yiru;QIAO Yuanhua(Beijing Key Laboratory of Internet Culture and Digital Dissemination Research,Beijing Information Science&Technology University,Beijing 100101,China;College of Applied Sciences,Beijing University of Technology,Beijing 100124,China)

机构地区：[1]北京信息科技大学网络文化与数字传播北京市重点实验室,北京100101 [2]北京工业大学应用数理学院,北京100124

出　　处：《北京信息科技大学学报（自然科学版）》2024年第1期37-41,共5页Journal of Beijing Information Science and Technology University

基　　金：北京市自然科学基金项目(4202025);天津市安监物联网技术企业重点实验室研究项目(VTJ-OT20230209-2);贵州省科技计划项目(ZK[2022]-012)。

摘　　要：基于极限学习机自编码器(extreme learning machine based autoencoder,ELM-AE)和误差反向传播(back propagation,BP)算法,针对ELM提出了一种改进的特征表示方法。首先,使用ELM-AE以无监督的方式学习紧凑的特征表示,即ELM-AE输出权重;其次,利用ELM-AE输出权重来初始化BP神经网络的输入权重,然后对BP网络进行监督训练;最后,用微调的BP网络输入权重初始化ELM的输入权重参数。在MNIST数据集上的实验结果表明,采用BP算法对ELM-AE学习的参数进行约束,可以得到更紧凑且具有判别性的特征表示,有助于提高ELM的性能。An improved feature representation method for extreme learning machine(ELM)was proposed based on extreme learning machine based autoencoder(ELM-AE)and error back propagation(BP)algorithms.Firstly,ELM-AE was used to learn compact feature representation i.e.ELM-AE output weights,in an unsupervised way.Secondly,ELM-AE output weights were used to initialize the input weights of the BP neural network,which was then trained by BP network in a supervised way.Finally,the input weight parameters of ELM were initialized by the input weight of the fine-tuned BP network.The experimental results on MNIST dataset show that using the BP algorithm to constrain the parameters of ELM-AE learning can result in a more compact and discriminative feature representation,which helps to improve the performance of ELM.

关 键 词：极限学习机自编码器 误差反向传播 极限学习机 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]