Most current artificial neural networks exist only within simulation programs running on conventional computers. Simulators can provide great flexibility, but require immensely powerful and costly hardware for even very small networks. In practice, this limits the size of artificial neural networks to a few thousand neurons (about the same as a house fly). If we are to find ways to mimic more advanced, human-like functions, we must find ways of implementing much larger networks, in a cost effective way.

Implementing an artificial neural network as a custom integrated circuit can provide a quantum leap in performance compared to a software simulator, at much lower unit cost. Each neuron can be treated as a separate processing unit, so all the neurons in a network can work simultaneously. Networks implemented this way can in principle operate many thousands of times faster than any simulator.

The larger the number of neurons and connections, the greater the complexity of the tasks that can be performed, but also the greater the silicon area required to implement the network on a chip. If the number of neurons and connections could be reduced, while still maintaining an acceptable level of accuracy and performance, more advanced networks could be implemented more cheaply.

This book discusses and quantifies the effectiveness of a number of algorithms for maximising the accuracy of a trained neural network given a reduced number of neurons and connections.

Contact Author: