Use of a structured problem domain to explore development of modularized neural networks

A technique has been devised and tested which allows separate training of neural network (NN) modules, each operating on a portion of the problem domain. A method for linking together the NN modules has been devised and shown to yield successful operation on the full problem domain. Methods in the graph theory literature known as connection matrix and reachability matrix were used to assist in both (i) decomposing the problem into subtasks and (ii) determining how to connect the NN modules that learn to perform the subtasks. The problem context used is a lattice of concept types underlying a knowledge base system.<
>