结合遗传算法的优化卷积神经网络学习方法

摘要

经典卷积神经网络学习方法采用最陡下降算法进行学习,学习性能受卷积层和全连接层的初始权重设置的影响较大.采用遗传算法生成多组初始权重,经过选择、交叉和变异操作得到最优权重;采用这些权重作为卷积神经网络的初始权重,其学习性能优于最陡下降算法随机选择的初始权重;采用遗传算法生成的多组权重训练多个卷积神经网络分类器,由其构建联合分类器进行分类,可进一步提高分类正确率.实验结果表明,与经典卷积神经网络方法以及常用的支持向量机、随机森林、后向传播神经网络和极速学习机相比,该方法的分类正确率更高.%The classic convolutional neural networks learning method uses steepest descent algorithm for learning.The learning performance is influenced largely by the initial weight setting of convolution layers and fully-connected layer.Genetic algorithm was used to generate multiple sets of initial weights,obtained by selection,crossover and mutation optimal weights.The optimal weights were obtained through selection,crossover and mutation operators.These weights were used as initial weights of convolutional neural network,which resulted in better performance than random selection of initial weights using steepest descent algorithm.Multiple convolutional neural network classifiers were trained using the multiple sets of initial weights generated using genetic algorithm,and combined classifier was then built for classification,which further improved the classification rate.Experimental results show that the proposed method has higher classification rate than the classical convolutional neural networks method and common methods such as support vector machine,random forest,back propagation neural network,and extreme learning machine.