Using Redundancy to Improve the Performance of Artificial Neural Networks*

For Artificial Neural Networks (ANNs) to be effective modelling tools, they must draw upon biological characteristics: One characteristic often overlooked in the design of ANNs is the replication, or redundancy, of processes within the brain. This paper examines the effects of redundancy on the performance of ANNs trained on either a pattern classification task (e.g. parity, encoder) or a function approximation task (e.g. forward kinematics). Results suggest that there is an optimal level of redundancy that increases the likelihood of network convergence while decreasing overall network processing time. ANNs with this level of redundancy consistently perform better than standard ANNs on pattern classification tasks. Furthermore, redundant ANNs trained on the function approximation task are more accurate in terms of overall system error than standard ANNs. These results imply that redundancy may be effectively used to increase the performance of ANNs, both in accuracy and speed.