多层去噪极限学习机

Multilayer denoising extreme learning machine

为提高极限学习机-自编码器(ELM-AE)所提取特征的有效性和鲁棒性,提升多层极限学习机(ML-ELM)性能,抑制噪声的影响,本文将极限学习机(ELM)与去噪自编码器(DAE)相结合,在ELM-AE中引入退化过程,提出了极限学习机-去噪自编码器(ELMDAE);并通过堆叠ELM-DAE,构建了多层去噪极限学习机算法(ML-D-ELM)。ML-DELM首先通过堆叠的ELM-DAE逐层提取具有高有效性、鲁棒性的抽象特征,而后运用ELM进行分类,完成由抽象特征到类别标签的映射。在多个常用数据集上的实验结果表明:与ELM、SAE-ELM、ML-ELM算法相比,无论是否存在噪声,ML-D-ELM的分类准确率都有明显上升。对于MNIST数据集,ML-D-ELM分类准确率可以达到98.81%。

In order to improve the effectiveness and robustness of the features extracted by the extreme learning machine autoencoder(ELM-AE),improve the performance of the multilayer extreme learning machine(ML-ELM) and suppress the influence of noise, an extreme learning machine denoising autoencoder(ELM-DAE)is proposed. This is based on combining the extreme learning machine(ELM)with denoising autoencoder(DAE) and introducing the corrupting process into ELM-AE,Then the multilayer denoising extreme learning machine(ML-D-ELM)is also proposed by stacking ELM-DAE. In ML-D-ELM,the highly effective and robust features are extracted by the stacked ELM-DAE,and then the classification is completed by ELM to map the abstract features to class label. Experimental results on some benchmark datasets show that ML-D-ELM can provide higher accuracy than ELM,SAE-ELM,and ML-ELM,even when considering noise. For MNIST,the accuracy of ML-D-ELM can reach98.81%·.