TRAINING REDUNDANT ARTIFICIAL NEURAL NETWORKS - IMPOSING BIOLOGY ON TECHNOLOGY

One biological principle that is often overlooked in the design of art ificial neural networks (ANNs) is redundancy. Redundancy is the replic ation of processes within the brain. This paper examines the effects o f redundancy on learning in ANNs when given either a function-approxim ation task or a pattern-classification task. The function-approximatio n task simulated a robotic arm reaching toward an object in two-dimens ional space, and the pattern-classification task was detecting parity. Results indicated that redundant ANNs learned the pattern-classificat ion problem much faster, and converge on a solution 100% of the time, whereas standard ANNs sometimes failed to learn the problem. Furthermo re, when overall network error is considered, redundant ANNs were sign ificantly more accurate than standard ANNs in performing the function- approximation task. These results are discussed in terms of the releva nce of redundancy to the performance of ANNs in general, and the relev ance of redundancy in biological systems in particular.