混合PSO优化卷积神经网络结构和参数

Optimizing Structure and Parameters of Convolutional Neural Networks Using Hybrid PSO

为了使卷积神经网络在非经验指导下自动寻得最优连接,并提高其参数优化效率,提出用粒子群优化卷积网络参数,并用离散粒子群优化卷积网络特征图之间连接结构的新方法。先使用粒子群优化所有权值,再采用离散粒子群优化降采样层和卷积层之间特征图连接结构。将该方法用于MNIST数据集和CIFAR-10数据集,实验结果表明,相比其他连接结构的卷积神经网络和其他识别方法,该方法可以有效实现网络结构及参数的优化,加速网络收敛并提高识别准确比。

In order to make convolutional neural network get optimal connection automatically withoutexperienced guidance and improve the optimizing effectiveness for parameters of convolutional neural network, anew method using both particle swarm optimization algorithm and discrete particle swarm optimization algorithmis proposed to optimize parameters and feature maps connecting structure of convolutional neural network. Theparticle swarm optimization is applied to optimize the weights of convolutional neural network at first, and then thediscrete particle swarm optimization is applied to optimize feature maps connections between sub-sampling layerand convolutional layer. The method is applied to MNIST database and CIFAR-10 database, compared toconvolutional neural networks of other connecting structures and other recognition methods, results shown that thismethod can optimize the parameters and structure of the network effectively, accelerate network convergence andimprove the recognition accuracy.