Research On Key Problems Of Transfer Learning In Deep Neural Networks

As the deep neural network(DNN)has been widely applied in various fileds,the problem of available data collection which requires a large amount of labor cost is also accompanied.In addition,the traditional machine learning algorithm assumes that the training data and the actual application data obey the same probability distribution,but the reality is usually different.With the continuous development of computer information technology,how to use the existing massive data and build a machine learning model in the case of a small amount of target domain data has become a current research focus.One of the most effective ways to solve this problem is Transfer Learning.It uses existing knowledge to solve different but related problems,so as to achieve knowledge transfer in related fields.Transfer learning relaxes the assumptions about the distribution of data,and it can find common features in source domain and target domain,so that knowledge can be shared among fields.This thesis studies the transfer learning of convolution neural networks,and proposes new convolution neural network transfer algorithms based on a large number of researchers.Here gives the main contents and innovations:(1)For the existing training DNN methods which can't get high accuracy with only a small amount of available data,it studies the more mature cross domain transfer learning theories,and then studies the transfer learning model of the deep neural network.In addition,it also studies the structure of the convolution neural network,the method of feature extraction and the change of the parameters of the convolution neural network in the transfer learning.(2)In order to solve the problem that the traditional convolutional neural network has a random mapping of the top-level classifier and the network parameters are all adjusted,two transfer learning algorithms used to improve the learning performance of the convolution neural network are studied.First,we starting from the structure of the convolution neural network,the task mapping algorithm is constructed through the combination of the similarity measurement algorithm and the minimum cost matching method.Secondly,starting from parameter training in the process of network transfer learning to constructi a k-selection transfer learning algorithm suitable for convolutional neural networks.We design experiments to verify the feasibility of the two algorithms.(3)We integrate task mapping and K selection transfer learning algorithm to form TMKT algorithm,and implement a simple transfer learning image classification system of small amount of data based on the browser platform.And design comparison test,the experimental results show that the convolutional neural network transfer learning algorithm proposed in this thesis has lower classification error rate.