Committee Networks: Many Heny Heads Are Better Than One

Abstract

Committee networks are based on the proverb that “many heads are better than one.” The neural networks in a committee called the committee members which have different capabilities will help each other and work together to solve a given problem. Basically, the outputs of the committee members will be combined through a certain fusion rule to produce the output of the committee network. This paper introduces the concept of committee networks along with a single neural network and reviews some of their current related literature. Here, the methods to form a committee network were classified and reviewed. This includes but not limited to the error decorrelation, linear combinations of neural networks, stacked generalization, and mixtures of local experts. All published works suggested that the committee networks perform better than a single neural network. Because of their performance, we conclude that committee networks should be used as a tool in several different fields including industrial and manufacturing engineering and be used as a replacement of a single neural network. If the computation time is a major concern however, using committee networks would not be appropriate. Further research directions in both theoretical and application are also provided in this paper.