Abstract: The Restricted Boltzmann Machine（RBM） is an effective feature extraction algorithm, inspired by the visual cortex sparse representation, people try to introduce the concept of sparse to the RBM, in order to learn to original data sparse representation, improve the capability of feature extraction. The Lorentz function is introduced to the RBM, as the RBM constrained sparse regularization is constructed based on Lorentz function of constrained sparse RBM model, referred to as LRBM（Lorentz function-based sparse constraints RBM）. The feature extraction performance of the model is evaluated visually, and the sparsity and classification rate are analyzed. Finally, multiple LRBM are superimposed, and the depth confidence network model based on LRBM is constructed and the performance of the depth network is analyzed. The experimental results show that the LRBM model can effectively extract the feature information of the data set, in the classification effect than the RBM average increase of about 2%, and improve the reliability of the target classification.

Key words: Restricted Boltzmann Machine（RBM）, sparse representation, feature extraction, Lorentz function-based sparse constraints RBM（LRBM）, target classification

摘要： 受限玻尔兹曼机（Restricted Boltzmann Machine，RBM）是一种有效的特征提取算法，受视觉皮层稀疏表示的启发，人们试图将稀疏这一概念引入到RBM中，以期学习到原始数据的稀疏表示，提高其特征提取性能。将Lorentz函数引入到RBM中，作为RBM的稀疏约束正则项，构建基于Lorentz函数的稀疏约束RBM模型，将其称之为LRBM模型。对该模型的特征提取性能进行了可视化评价，同时对稀疏度和分类率进行了实验分析；最后将多个LRBM叠加，构造基于LRBM的深度置信网模型并分析该深度网络的性能。实验表明，LRBM模型有效地提取了数据集中的特征信息，在分类效果上较RBM平均提高了2%左右，增强了目标分类的可靠性。

关键词: 受限玻尔兹曼机（RBM）, 稀疏表示, 特征提取, LRBM模型, 目标分类