A Distributed Learning Algorithm for RBF Neural Networks

Abstract

Training a radial basis function (RBF) neural network on a single processor is usually challenging due to the limited computation and storage sources, especially for data with large and multi-dimensional features. In addition, in real applications, large-scale data may be collected in a distributed manner, which also makes it difficult to handle the data only with a single processor. To address these issues, we propose a distributed learning algorithm for RBF neural networks. In this algorithm, RBF neural networks can be trained in parallel using multiple processors. Specifically, the large-scale training data is divided into groups and each processor is associated with only one group. By introducing a shared output weight vector, training can be carried out simultaneously on different processors. The formulated optimization problem is addressed with alternating direction method of multipliers (ADMM) framework. Simulation results demonstrate the effectiveness of the proposed algorithm.