基于去噪自编码器的镜像极限学习机设计

Mirror Extreme Learning Machine Based on Denoising Autoencoder

针对极限学习机在高维度、含噪声数据集中需要大量隐含层节点来保证分类性能的问题,设计了镜像极限学习机.该算法使用伪逆法确定输入权值,随机生成输出权值和偏置,在对数据进行分类时,它仅需极少的隐含层节点.为了提升镜像极限学习机的分类性能和抗噪性,将它与去噪自编码器相结合.利用去噪自编码器对输入数据进行特征提取,并将提取到的特征作为镜像极限学习机的输入数据,再进行网络训练.在无噪和含噪声的MNIST,Fashion MNIST,Rectangles和Convex数据集中,将基于去噪自编码器的镜像极限学习机与ELM,PCA-ELM,SAA-2和DAE-ELM作对比实验,结果表明,基于去噪自编码器的镜像极限学习机的综合性能最优,用于分类的网络隐含层节点数最少.

A mirror extreme learning machine(MELM)is designed aiming at the problem that a large number of hidden layer nodes are needed to guarantee the classification performance of the extreme learning machine(ELM)in the high-dimensional and noisy data set.The MELM uses the pseudo-inverse method to determine the input weights,and randomly generates output weights and biases.When classifying data,it requires only a few hidden layer nodes.In order to improve the classification performance and noise immunity of the MELM,it is combined with a denoising autoencoder(DAE).The DAE is used to extract features from the input data,and the extracted features are used as the input data of the MELM,and finally the network training is triained.In the noiseless and noisy MNIST,Fashion MNIST,Rectangles and Convex data sets,the MELM based on the DAE is compared with the algorithms ELM,PCA-ELM,SAA-2 and DAE-ELM.The experimental results show the superiority of the algorithm:the MELM based on the DAE uses the least number of hidden nodes in the network for classification,and the overall performance is the best.