深度学习算法的原理及应用  

Principles and practices of deep learning

作　　者：胡侯立 魏维[1] 胡蒙娜[2] 

机构地区：[1]西安通信学院,西安710106 [2]陕西师范大学教育学院,西安710062

出　　处：《信息技术》2015年第2期175-177,共3页Information Technology

摘　　要：深度学习作为一种新兴的多层神经网络降维算法,通过组建含有多个隐层的神经网络深层模型,对输入的高维数据逐层提取特征,以发现数据的低维嵌套结构,形成更加抽象有效的高层表示。从深度学习算法的基本原理入手,较为详细地讲解了深度学习算法的单层网络基本结构受限波尔兹曼机及其训练过程。最后通过自动编码机举例说明深度学习技术应用于手写数字识别所带来的性能提升,并对深度学习技术做了简单总结。It is possible to find some effective low-dimensional codes in original high-dimensional data. Deep learning as a new developmental multilayer neural network algorithm possess of many hidden layers in neural network. This algorithm take layer-wise way to extract features form the high-dimensional dada composing more abstract and effective high-level expression, and find the low-dimensional feature of original data. Autoencoder is made of many restricted boltzmann machines, by pertaining network model to initialize the weights use Contrastive Divergence algorithms, and by minimizing the discrepancy between the original data and its reconstruction. The experiment show autoencoder is a efficient model in handwritin~ recoznition field.

关 键 词：深度学习 神经网络算法 受限波尔兹曼机 自动编码机 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]