Construction of the Convolutional neural network Based on the increase optimizers provided and atomic layers provided

Abstract

In order to solve the problem of low accuracy of traditional neural network code test set and verification set, we use VGG network structure to deal with it. Its main feature is that it has a simple structure and replaces all convolution kernels with 3 × 3. Compared with the pooled cores of Alexnet, VGG uses 2 × 2 pooled cores. Number of parameters is large, and most of the parameters are concentrated in the full connection layer. Proper network initialization and use of batch normalization layer are very important for training deep network. In Alexnet, there is only one convolution in each convolution layer, and the convolution kernel is 7*7. In VGGnet, each convolution layer contains 2 ~4 convolution operations. The most obvious improvement of VGGnet is to reduce the size of convolution kernel and increase the number of convolution layers. Based on the traditional convolutional neural network, we train a better classifier to realize the recognition and classification of cifar-10 data sets. The accuracy of verification set is improved from 52.9% to 70.3%, and the accuracy of test set is improved from 54.7% to 71.5%.