一种用于降维和盲源分离的主独立元神经网络

Principal Independent Component Network for Dimensionality Reduction and Blind Source Separation

经典主元分析和主元神经网络常以主元所能提取总的系统方差来确定主元数目 ,这隐含假设系统数据是高斯分布 ,所提取的主元之间相互无关 ,但不一定相互独立 ,从而难以实现非高斯系统数据的降维和信源分量。针对非高斯随机系统数据的降维和信源分离问题 ,提出一种基于二阶 Renyi近似熵指标的主独立元神经网络 ,并给出熵的近似计算方法及相应的梯度学习算法。仿真实验证明 ,该主独立元网络不仅能对数据降维压缩 ,还能有效地分离出普通主元分析法所不能提取的独立信源信息。

Principal component analysis and principal component neural network generally use the index of the total variance interpreted by principal components to choose the adequate number of principal component. These approaches implicitly suppose that the system data are Gaussian distribution and may be inappropriate for the dimensionality reduction of the non-Gaussian data. Considering the dimensionality reduction and the blind source separation of the mixture data from non-Gaussian stochastic systems, a principal independent neural network based on second order Renyi entropy criterion is proposed. An approximation method for the computation of the Renyi entropy criterion and the corresponding gradient learning algorithm are given. Simulation example shows the effectiveness of the approach for the dimensionality reduction and its advantages of the blind source separation over general principle component analysis.