摘要
近年来,基于深度学习的自编码神经网络是数据降维问题研究的热点,数据降维能够有效地消除无关和冗余信息,提高学习数据内在特征的效率.研究了在原始数据预处理时加入噪声,可训练出对输入信息更加鲁棒的表达,从而提升自编码神经网络模型对输入数据的泛化能力.提出了一种稀疏降噪自编码神经网络(Sparse De-noising Auto-Encoder,SDAE),基于稀疏性的思想,对降噪自编码神经网络加以改进,使得抽象出的特征稀疏表示,更有效的用于数据分类.实验结果表明稀疏降噪自编码神经网络(SDAE)分类准确率要优于传统的自编码神经网络及降噪自编码神经网络.
In recent years, the study about auto-encoder neural network based on deep learning has been a hot topic in research of data dimension reduction, which can eliminate irrelevant and redundant information effectively and im-prove the efficiency of the inherent characteristics of the learning data. More robust expression for input data can be trained through adding noise at the raw data preprocessing, which thereby enhances the generalization of auto-encod-er neural network model for input data. Sparse De-noising Auto-Encoder(SDAE)was proposed. De-noising au-to-encoder neural networks were enhanced based on the idea of sparsity which enables abstract features of sparse representation to become more effective for data classification. Experimental results show that classification accuracy of SDAE is better than that of traditional auto-encoder neural network and de-noising auto-encoder neural network.
出处
《内蒙古民族大学学报（自然科学版）》
2016年第1期21-25,93,共6页
Journal of Inner Mongolia Minzu University：Natural Sciences
基金
国家自然科学基金资助项目(61163034
61373067)
内蒙古自治区自然科学基金资助项目(2013MS0910
2013MS0911)
关键词
数据降维
降噪
稀疏
稀疏降噪自编码神经网络
Dimension reduction
De-noise
Sparse
Sparse De-noising Auto-Encoder neural network
