摘要
针对高维连续数据的降维问题,提出一种新的非线性降维方法,称为连续自编码(Continuous autoencoder,C-autoencoder)神经网络,该方法采用限制玻耳兹曼机的连续形式(Continuous restricted Boltzmann machine,CRBM)的网络结构,通过训练具有多个中间层的双向深层神经网络将高维连续数据转换成低维嵌套并继而重构高维连续数据。这种连续自编码网络提供了高维连续数据空间和低维嵌套结构的双向映射,有效解决了大多数非线性降维方法所不具备的逆向映射问题,特别适用于高维连续数据的降维和重构。将C-autoencoder用于连续帧图像的实验表明,C-autoencoder不仅能发现嵌入在高维连续帧图像中的非线性低维结构,也能有效地从低维结构中恢复原始高维图像数据,而且还能对连续帧图像有效地进行内插重构。
To solve the problem of the dimensionality reduction for high-dimensional continuous data,a novel nonlinear dimensionality reduction method,called the ＂continuous autoencoder＂,is proposed.The method uses continuous restricted Boltzmann machine（CRBM） network structure and converts high-dimensional continuous data to low-dimensional codes by training a neural network with multiple hidden layers.In particular,the ＂continuous autoencoder＂ provides a bi-directional mapping between the high-dimensional continuous data space and the low-dimensional manifold space,thus overcoming the inherited deficiency of most nonlinear dimensionality reduction methods without an inverse mapping.Experiments on sequential images show that the ＂continuous autoencoder＂ network can find the embedded manifold with high-dimensional sequential image and reconstruct the original high-dimensional sequential image from the low-dimensional structure.It can interpolate reconstruction sequential image.
出处
《数据采集与处理》
CSCD
北大核心
2010年第3期318-323,共6页
Journal of Data Acquisition and Processing
关键词
高维连续数据
降维
连续自编码网络
内插重构
high-dimensional continuous data
dimensionality reduction
continuous autoencoder network
interpolation reconstruction
