一种基于集成学习与聚类聚合的联邦学习预训练方法OA

Data generated in different data centers are becoming more and more important because of its resource properties for deep learning.However,due to realistic communication conditions and security,data in data centers cannot be transmitted to each other.To make use of this data,a machine learning algorithm called federated learning is developed.Federated learning en-ables multiple customers in different geographic locations to collaborate on learning machine learning models while keeping all of their data on the device.However,the traditional federated learning algorithm does not consider the influence of the initial optimiza-tion point on the final generalization performance.In the actual applications,the selected random initial optimization points do not significantly improve the optimization performance.Therefore,in order to solve this problem,a pre-training method called FedPre,which can effectively generate initial points with better performance for federated learning,is proposed.FedPre uses ensemble learn-ing to effectively improve the training performance of the initial point by creating a larger ensemble model to learn features more widely in the dataset.Finally,experiments and performance analysis are performed on the FASHION-MNIST and CIFAR-10 datas-ets.In the experiments,several different federated learning algorithms are used to train the initial points obtained by FedPre.The ex-perimental results show that the initial points obtained by FedPre algorithm can significantly improve the generalization ability of fed-eration learning algorithms.