基于PCA的GHA算法在BP网中的融合应用  

Mixed Application of GHA Based on PCA in BP Neural Network

作　　者：范燕[1] 吴小俊[2] 邵长斌[1] 宋晓宁[2] 

机构地区：[1]江苏科技大学计算机科学与工程学院,镇江212003 [2]江南大学物联网工程学院,无锡214122

出　　处：《计算机科学》2015年第12期302-306,共5页Computer Science

基　　金：国家自然科学基金(61100116;61471182;61572242);江苏省自然科学基金(BK2012700;BK20130473);江苏省高校自然科学基金(13KJB5200003);四川省人工智能重点实验室基金(2012RZY02)资助

摘　　要：针对传统特征提取方法和BP神经网络相结合而存在的缺点,提出一种新的分类器模型"PCABP网"。首先利用PCA特征向量来初始化PCABP网的初始层权值矩阵,由此新模型的初始层起到取代PCA进行特征提取的作用。其次在训练过程中通过GHA和GD算法对初始层投影权值矩阵进行动态调节来优化特征向量。该方法从源头样本来优化"类别分离"和"特征提取",找到对样本降维和分类的最佳契合点,以此来替代传统模式识别中"首先单独特征提取,其次利用分类器分类"的方式。在FERET人脸库上的实验结果验证了该方法的有效性。In view of the defects resulting from the combination of traditional method of feature extraction and BP neural network, this paper presented a new classification model ＂PCABP network＂. Firstly, the PCA eigenveetor is used to initialize the initial-layer weight matrix of the PCABP network, thus, the initial-layer of the new classification model ＂PCABP Network＂ replaces the role of PCA in the function of feature extraction. Secondly, in the training process, with the application of GHA and GD algorithm, dynamic adjustment to the projection weights matrix of the initial layer has been achieved,and accordingly, the PCA eigenvector has been optimized. This method optimizes ＂category separation＂ and ＂feature extraction＂ from source samples, finds out the best connection point between sample dimension reduction and classification, and replaces the traditional recognition pattern ＂firstly separate {eature extraction, then classification by the use of classifier＂. The experiment based on FERET face library verifies the effectiveness of this method.

关 键 词：前馈神经网络 主成分分析 广义赫布算法 弹性梯度下降算法 性别分类 融合神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]