基于RVFL神经网络的集成深度学习  

Ensemble Deep Learning Based on RVFL Neural Network

作　　者：郭金成 魏霖静[1] GUO Jincheng;WEI Linjing(School of Information Science and Technology,Gansu Agricultural University,Lanzhou 730070)

机构地区：[1]甘肃农业大学信息科学技术学院,兰州730070

出　　处：《计算机与数字工程》2025年第3期786-794,共9页Computer & Digital Engineering

基　　金：2020年甘肃农业大学研究生教育研究项目(编号:2020-19);2021年度兰州市人才创新创业项目(编号:2021-RC-47);2021年教育部产学研合作协同育人项目(编号:202102326036)资助。

摘　　要：论文提出了基于随机神经网络的深度学习框架。受随机向量函数链接(Random Vector Functional Link,RVFL)网络原理的启发,提出一种了具有堆叠层的深度RVFL网络(dRVFL)。该网络能够在提取多个隐藏层特征的同时充当加权网络,从而为直接链接获得的原始特征以及隐藏层的特征提供权重。其次,还提出了一种集成学习与深度学习相结合的集成深度RVFL网络(edRVFL)。相较于传统集成方法,edRVFL仅通过训练一次单个dRVFL网络即可。该网络中封闭形式的解决方案极大程度地缩短了非迭代训练时间。同时,这两种网络的框架可与任何RVFL变体共同使用。最后,将论文提出的网络与9种不同的深度前馈神经网络在46个表格UCI分类数据集和12个稀疏数据集上进行比较实验,结果表明论文提出的网络性能更优。This paper proposes a deep learning framework based on stochastic neural network.Inspired by the principle of Random Vector Functional Link(RVFL)network,a deep RVFL network with stacked layers(dRVFL)is proposed.The network can be used as a weighted network while extracting the features of multiple hidden layers so as to provide weights for the original features obtained by direct links and the features of hidden layers.Secondly,an integrated deep RVFL network(edRVFL)combining integrated learning and deep learning is proposed.Compared to traditional integration methods,edRVFL only trains a single dRVFL network once.The closed-form solution in the network greatly reduces the non-iterative training time.At the same time,the frameworks for both networks can be used with any RVFL variant.Finally,the proposed network is compared with nine different deep feedforward neural networks on 46 tabular UCI classification datasets and 12 sparse datasets,and the results show that the proposed network has better performance.

关 键 词：RVFL 深度学习 集成学习 深度前馈神经网络 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]