A learning scheme of neural networks which improves accuracy and speed of convergence using redundant and diversified network structures

Abstract

The author proposes a learning scheme which compensates for the incomplete result of learning using redundant internal coding of the required input-output relation and some plans to diversify inner subnetwork structures. He applies this scheme to a character recognition problem and experimentally shows that this approach gives more accurate learning results and faster convergence as well as more efficient hardware constitutions than the traditional approach. Specifically, computer simulations are presented which shows that the proposed approach is superior to the traditional approach using the so-called grandmother cell representation scheme.<>