Performence analysis of neural network with improved weight training process
Yifeng Zhao, Weimin Lang, Bin Li · 2019 IEEE 3rd Information Technology, Networking, Electronic and Automation Control Conference (ITNEC) · 2019
The traditional convolutional neural network usually initializes the weights of all network layers at one time before network training, and then updates the weights of the network by back-propagation algorithm to improve the accuracy of the network during network training. However, with the increase of network depth, the computational cost of this method will increase dramatically and the test accuracy will be affected. In order to solve this problem, a method of gradually reinitializing the weights of each layer is proposed, that is, after a certain training period, the weight of the previous layer is determined and remain unchanged, then initialize the weights of all subsequent layers, repeat this step until the weights of all layers are determined. In order to verify the performance of the method, a series of experiments were carried out on the CIFAR10 dataset. The results show that the accuracy of the network is improved by 9% and the training time is reduced by 29%. It shows that the method can improve the accuracy of the network and reduce the training time.