Automatic task decomposition in modular networks by structural learning with forgetting

It is well known that large-scale homogeneous networks suffer from serious difficulties: the scale problem and the local minima problem. A promising approach to alleviating these is the use of modular networks. There have been various studies on modular networks. A competitive modular connectionist architecture by Jacobs et al. (1991) accomplishes automatic task decomposition. Its architecture consists of a set of expert networks and a gating network. We propose a simpler method for the task decomposition in modular networks using the structural learning with forgetting. It has the ability to generate a skeletal structure due to forgetting. Because of this property, automatic task decomposition without a coordinating mechanism becomes possible. As a result, each subtask is undertaken by one of the modules. Its application to the recognition of geometrical figures fully demonstrates its effectiveness in the task decomposition.