Optimization Of Convolutional Neural Network Structure Based On Particle Swarm Optimization

Convolutional Neural Network(CNN)is widely applied in various fields.The performance of trained neural network classifier depends on its structure,however it is very difficult to find an optimal CNN structure.This problem seems to be a major obstacle to CNN application.First of all,the structure of CNN is complex and diverse.The parameters that determine CNN structure are almost hyper-parameters which can't be obtained through training of network.So they need to be manually set and combined.Unfortunately,the existing methods depend on experience.Secondly,the initial weights of CNN training have a certain influence on the learning of CNN.A good CNN structure is robust and can weaken the influence of initial weights.Finally,for different problems,the CNN structure is different and the CNN structure needs to be re-selected.This paper conducts the following research on the existing problems of CNN structure optimization.(1)The structure of CNN and the existing algorithms for optimizing CNN are studied systematically.By analyzing the algorithm principle and experimental results,the importance of CNN structure optimization is discussed.(2)Based on the existing algorithms which use Particle Swarm Optimization(PSO)to optimize the CNN structure,a new improved Particle Swarm Optimization Convolutional Neural Network structure algorithm(IPSOCNN)is proposed.This algorithm increases the number of hyper-parameters and enriches the diversity of CNN structure.The IPSOCNN algorithm automates the optimization of the CNN structure and solves the problem of manual selection uncertainty.Experiments on the MNIST dataset and AR dataset show that,the IPSOCNN algorithm has a higher correct rate than that of the other optimal CNN structures and the existing optimized CNN algorithms.(3)After making the experimental analysis,the IPSOCNN algorithm still has further optimization space.Since the Quantum Particle Swarm Optimization(QPSO)is better than the PSO algorithm in performance and its parameters are automatically selected by learning,this paper proposes a new method called Optimization of Convolutional Neural Network Structure Based on Quantum Particle Swarm Optimization(QPSOCNN).The results of experiments indicate that the QPSOCNN algorithm has higher accuracy than that of the IPSOCNN algorithm.