Journal of Frontiers of Computer Science and Technology ›› 2018, Vol. 12 ›› Issue (5): 820-827.DOI: 10.3778/j.issn.1673-9418.1702037

Previous Articles     Next Articles

ELM Optimized Deep Autoencoder Classification Algorithm

XU Yi, DONG Qing+, DAI Xin, SONG Wei   

  1. School of Internet of Things Engineering, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Online:2018-05-01 Published:2018-05-07


徐    毅,董    晴+,戴    鑫,宋    威   

  1. 江南大学 物联网工程学院,江苏 无锡 214122

Abstract:  For the autoencoder neural networks having long training time, this paper puts forward a kind of improved deep autoencoder neural network. Firstly, this paper uses extreme learning machine (ELM) as an autoencoder block and constructs a multilayer autoencoder neural network, to improve the classification accuracy. Using ELM can avoid the iterative process and reduce the training time of network. Secondly, this paper adds the label tag in the output layer nodes, and expects the actual output with the sample tag, making the unsupervised learning to be a supervised learning and achieving the classification training in the process of deep learning. To verify the validity of the proposed method, this paper tests in the multiple UCI datasets. The experimental results show that the accuracy is good and the training speed is improved, compared with other autoencoder networks and radial basis function (RBF) neural network.

Key words: deep neural network, extreme learning machine, autoencoder, classification

摘要: 针对自编码神经网络训练时间长的问题,提出了一种改进的深度自编码神经网络算法。首先利用极限学习机(extreme learning machine,ELM)作为自编码块,构建多层自编码神经网络,以提高分类准确率。采用ELM能避免大量的迭代过程,减少网络训练时间。其次为实现分类,在各输出层中加入标签节点,对实际输出与各样本的期望标签进行比对,使原始的自编码无监督学习转化为监督学习过程,从而在深度学习的过程中实现分类训练。为验证该方法的有效性,在多个UCI数据集中进行广泛的测试。实验结果表明,与其他自编码网络和RBF(radial basis function)神经网络相比,该方法取得了良好的分类准确率,并且有效提高了训练速度。

关键词: 深度神经网络, 极限学习机, 自编码, 分类