Robust Expectation Maximization Learning Algorithm for Mixture of Experts

Abstract

The Mixture of Experts (ME) model is a type of modular artificial neural network (MANN) specially suitable when the search space is stratified and whose architecture is composed by different kinds of networks which compete to learn several aspects of a complex problem. Training a ME architecture can be treated as a maximum likelihood estimation problem, where the Expectation Maximization (EM) algorithm decouples the estimation process in a manner that fits well with the modular structure of the ME architecture. However, the learning process relies on the data and so is the performance. When the data is exposed to outliers, the model is affected by being sensible to these deviations obtaining a poor performance as it is shown in this work. This paper proposes a Robust Expectation Maximization algorithm for learning a ME model (REM-ME) based on M-estimators. We show empirically that the REM-ME for these architectures prevents performance deterioration due to outliers and yields significantly faster convergence than other approaches.

This work was supported in part by Research Grant Fondecyt 1010101 and 7010101, in part by Research Grant CHL-99/023 from the German Ministry of Education and Research (BMBF) and in part by Research Grant DGIP-UTFSM and in part by the Intership grant CONICYT-INRIA