Development of a learning system for convolutional neural network

Abstract

Traditional learning system of convolutional neural network (CNN) is based on gradient descent method and back propagation. Effective though it is, we still keep seeking new learning system to make possible progress. For this purpose, we try to utilize incremental learning algorithm which is shown effective on the approximation of robot kinematic model and forward thinking framework as alternatives. This dissertation is mainly about the preliminary work we do before combining these two algorithms together. We investigate the performance of incremental learning algorithm for a single hidden layer feed forward network and the fully connected part of CNN, with comparison to traditional learning algorithm. We also verify the effectiveness of the forward thinking framework in CNN. Furthermore, the effect of the number of layers of shallow network in forward thinking framework is investigated.