Abstract

The Hierarchical mixture of experts(HME) architecture is a powerful tree structured architecture for supervised learning. In this paper, an efficient one-pass algorithm to solve the M-step of the EM iterations while training the HME network to perform classification tasks, is first described. This substantially reduces the training time compared to using the IRLS method to solve the M-step. Further, a pre-processing stage is proposed, consisting of radial basis function kernels, aimed at reducing the tree height of the HME network. Alternatively, employment of a localized form of gating network is suggested to reduce the tree height. Shorter HME trees, with much fewer network parameters, are significantly faster to train. Simulation results are presented on a real life data set. 1. INTRODUCTION The Hierarchical mixture of experts(HME) of Jordan and Jacobs [1] is a tree structured architecture which works on the principle of divide and conquer. The model employs probabilistic methods t...

Citations