Multimodule Neural Network for Complex and Dynamic Associations

Abstract

A multimodule neural network model for flexible associative memory is proposed. This multimodule network can deal with complicated many-to-many association and cope with complex input-output relations more flexibly than conventional associative memories. Each module is characterized as a hybrid architecture of an autoassociative recurrent network and a heteroassociative feedforward network. A module memorizes partial patterns of complete patterns individually, and the connections between these modules learn interdependencies between them. Since each module is constructed such that it can recall multiple patterns sequentially and automatically from a key input, the network can also make many-to-many recollections.