| （英） |
To improve the performance of machine learning models, it is necessary to train them using large datasets. However, it is difficult in some cases to obtain a large dataset due to privacy issues. To address this problem, federated learning (FL) has been proposed, in which data owners train their models individually and share the model parameters to train the model without collecting the data. We have proposed a serverless federated learning method for IoT devices and sensor networks in [1], which reduces communication costs and enables learning in a heterogeneous model environment by using knowledge distillation (KD). In this paper, we propose to use the alternating direction method of multipliers (ADMM) to achieve faster convergence than consensus-based optimization, which is utilized in [1]. The proposed method emulates the ADMM in function spaces to optimize models, and this procedure is realized by updating the model parameters by distillation. We also compare the convergence speed of the proposed method with that of existing methods through simulations. |