Accuracy Improvement Based on Classic Neural Network: Voting, Restarting and Quantization

Abstract Since AlphaGo beat the world Go champion in 2016, which attracted wide attention, neural network has become more and popular recent years, people’s research on it gradually improved been applied different fields. Today, artificial intelligence machine learning have an essential part of modern society intelligent systems. We do image recognition, speech visual learning, closely related to learning. However, unsatisfactory training results or even failure is always encountered. Therefore, imperative improve accuracy results. In this paper, processing, MNIST, other classical models will be used set parameters better through voting, quantization, restart, methods. The aiming find relationship between restart numbers process total extent improvement. At same time, several algorithms utilizing these are compared selected. Finally, conclusion drawn that made with a convolutional network, less profit improvement we gain from restarting process.