Locally versus globally interlayer-connected feed-forward neural networks: a performance comparison
John D. Provence, S. Naganathan · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 1990
During the past few years, artificial neural networks have been applied to problems such as pattern matching, associative memory recall, and optimization problems. A property common to all the network structures which have been proposed and studied is global interconnectivity among the neural processing nodes. While global interconnectivity presents few problems for implementing simulations of networks, it is a major factor which prohibits implementation using VLSI and or optical technology. In this paper, we present the results of a study which compares the performance of globally interlayer connected multi-layer neural networks with that of networks which employ local connections. The locally connected neural network contains multiple hidden layers and each neural processing node in a given layer connects to at most its three nearest neighbor processing nodes in the higher layer. The advantages of a locally connected network are reduced interconnection complexity, reduced I/O requirements for each neural processor node, and faster processing at each neural processing node. The locally connected and globally connected networks are compared with respect to training iterations, number of neural processor nodes needed, number of hidden layers needed, and error rate performance for a number of different problems.