Application Of Deep Learning Algorithm For Multilayer Convolution Neural Networks

Deep Learning(DL)has been a hot research in the field of artificial intelligence.It learns the essential characteristics of massive data through multi-hidden nonlinear unit,and realizes the approximation of high-order abstract function to improve the accuracy of classification or prediction.Compared with shallow learning,the structure of the depth learning is more similar to the cerebral cortex,the input data from input layer is processed hierarchically by each hidden layer,so the ability to express complex functions and computing is stronger.Convolution neural network(CNN)is the most basic and effective algorithm in depth learning.The local experience,weight sharing and down-sampling structure greatly reduce the complexity of the network model,reducing the number of training weights.In the field of image recognition,the unique structure of CNN has a high degree of distortion in translation,scaling or other deformations.Therefore,it is of great significance to apply the depth learning technology to the field of image recognition.Based on the research at home and abroad,this paper summarized and introduced the basic structure and algorithm of multi-layer feedforward network and CNN,and deeply studied the parameter design of CNN and the optimization problem of stochastic gradient descent(SGD)learning algorithm.The paper applied the Python language to the multi-layer CNN model design on the Theano development platform,and applied the improved SGD learning algorithm to handwritten numeral recognition and face recognition.In this paper,two kinds of network models,such as convolution neural network and multi-layer feedforward network,have been designed.The error sensitivity and learning rate of different hidden layers have been analyzed.The influence of different learning rate on the convergence of convolution neural network has been studied.Then,the paper focused on the improvement of Stochastic Gradient Descent(SGD).By optimizing the learning rate parameters of the network and testing on the ORL data set,the algorithm was validated on the face recognition problem feasibility.Finally,the influence of convolution kernel parameters on the recognition rate and training rate of convolution neural network was studied by analyzing the experimental results of three different sets of network models in face recognition test set.In this paper,we studied the network structure and parameter optimization in different databases,and summarized some practical rules of depth learning in the application of image recognition and classification,and proved that the convolution neural network could deal with complex image classification problem.The recognition error rate(0.84%)of the network model with improved learning rate algorithm was lower than that of the multi-layer feedforward network(2.52%)in the MNIST data set,which was better than that of the model(6.25%)in the literature.At the same time,in the ORL database,the network with improved algorithm was superior to the traditional network and the literature in terms of recognition rate and training speed.Experiments showed that the algorithm had good practical performance and had a good guiding effect on solving practical engineering problems.