City Research Online

Abstract

A novel method of integrating multiple neural networks into one large network via a process referred to as a neural network linking process is proposed.

Neural networks are commonly trained to solve a specific problem for an encapsulated problem domain. A single network can undertake simple classification or generalisation problems. Dividing them into sub-problems, which in turn are solved by a sub-network, can disentangle more complicated classification or generalisation problems. A controller generally combines sub-network results. A controller can be, for instance, a gating network, voting system or a mathematical combiner. In each case, every sub-network is used as a separate unit and is not interconnected to any other sub-network.

However, with the linking process a novel method for linking trained sub-networks into one large network by maintaining the knowledge of each individual sub-network is introduced. Furthermore, the linked network will utilize a stimulus process in order to distinguish the type of sub-problem to be solved, by largely retaining the accuracy of the sub-network, as well as being one step closer to the biological reality.

Publication Type:	Thesis (Doctoral)
Subjects:	Q Science > QA Mathematics Q Science > QH Natural history > QH301 Biology
Departments:	School of Science & Technology > Mathematics School of Science & Technology > School of Science & Technology Doctoral Theses Doctoral Theses

Export

Downloads