Sequential Bayesian Learning for Modular Neural Networks

In this paper, we present a distributed computing method, namely Sequential Bayesian Learning for modular neural networks. The method is based on the idea of sequential Bayesian decision analysis to gradually improving the decision accuracy by collecting more information derived from a series of experiments and determine the combination weights of each sub-network. One of the advantages of this method is it emulates humans' problems processing mode effectively and makes uses of old information while new data information is acquired at each stage. The results of experiments on eight regression problems show that the method is superior to simple averaging on those hard-to-learn problems.