手写数字识别中组合式神经网络的构建方法

Approach for Constructing Neural Network Ensemble Applied to Handwritten Digit Recognition

将聚类技术和遗传算法相结合,提出一种基于相似度传播算法和遗传算法的神经网络集成方法应用于手写数字识别问题.先分别利用主成分分析和Fisher线性判别分析对数据集进行特征提取,得到两类特征数据集,再利用Bagging方法分别为这两类特征数据集训练简单的BP神经网络,然后采用相似度传播算法对这些BP神经网络进行聚类,找到作为类簇中心的网络(中心网络),最后利用遗传算法对所有中心网络的权值进行训练,将中心网络进行加权线性集成作为最终分类器.在标准手写数字数据集MNIST上进行测试的实验结果表明,该方法的识别率优于单个神经网络的识别率,并兼顾了分类效率.

The authors proposed a method for constructing neural network ensemble based on affinity propagation and genetic algorithm, which can be applied to handwritten digit recognition. Firstly, we extracted features of samples from the MNIST handwritten digit database by PCA （ Principle Component Analysis） and Fisher LDA （Linear Discriminant Analysis） respectively. Secondly, we adopted the bagging method to train some BP neural networks as candidate networks using the extracted PCA and LDA feature sets. Thirdly, affinity propagation algorithm was used to group candidate networks and find exemplars in each cluster. Finally, we optimized these weights for exemplars using genetic algorithm and integrate exemplars with the optimized weights to obtain the final classifier. Experimental results on the MNIST handwritten digit database prove that the accuracy of the algorithm proposed in this paper is far higher than that of individual network while its efficiency is acceptable.