Design And Implementation Of Edge Transfer Learning In Edge-cloud Collaborative Deep Learning Model

With the development of artificial intelligence technology and the popularization of smart devices,edge-cloud collaborative deep learning technology based on edge computing and cloud computing has been paid great attention.Edge-cloud collaborative deep learning deploy deep learning model in edge devices and cloud server in a distributed manner,operating in a collaborative manner between edge model and cloud model,fully utilizing the advantages of edge computing and cloud computing.However,in reality,the deployment scenarios of edge devices often change due to natural factors.The edge model trained in a single scenario is difficult to cope with the changing scenarios,so the model performance will fluctuate greatly with the changes of the scenario.Transfer learning can apply the knowledge of the old scenario data learned by the edge model to the new scenario,so that the edge model has a better generalization ability for different scenarios.This paper takes the Branchy Net deployed on edge devices as the research object,and proposes different transfer learning methods to solve the problem of performance fluctuations due to scenario changes.To solve model parameters transfer problem of Branchy Net,this paper combines three typical transfer learning methods,such as global distribution alignment,subspace distribution alignment,and local distribution alignment,with Branchu Net to perform transfer learning on the model parameters.And based on the ideas of these methods,a model parameters transfer method using the characteristics of early exiting is proposed.The proposed method labels the simple samples and difficult samples of the source domain and the target domain with a clustering method,and uses the maximum mean difference to align them,which improves reasoning speed and accuracy of Branchy Net.This paper uses a standard public transfer learning dataset to experimentally verify the transfer learning method of Branchy Net parameters,and the results show that the proposed method is effective.To solve exit threshold transfer problem of Branchy Net,this paper proposes a feature vector transfer method based on a supervised decision maker,which transfers the feature vector generated by the source domain and target domain samples,and then uses a supervised machine learning model to make the feature vector exit decision.This paper also proposes an exit threshold transfer method based on an unsupervised decision maker,which transfers the predicted entropy values of the source domain and target domain samples,and then uses an unsupervised machine learning model to segment the entropy values to determine whether samples exit.This paper uses standard transfer learning dataset to test two different methods,and the results show that the entropy transfer method is more effective than the feature vector transfer method.