A multi-computer neural network applied to machine-vision

The backpropagation neural network is recognised to have a convergence rate which is slower than desired. Transputer systems are attractive platforms for the implementation of neural networks, offering the potential for achieving faster convergence through increased processing power. However, multiple-transputer implementations of the backpropagation algorithm which are found in the literature offer an improvement in performance which is less than would be anticipated due to the inherent high communications overhead. This paper describes the class-distributed (C-D) network, a new method of implementing a modified backpropagation algorithm on a multiple-transputer system to form a multi-computer classifier. The communications requirement for this new network is minimal and the convergence rate is superior to that of comparable methods. The performance of the C-D network is evaluated in a machine vision application where shape is used for object identification