The suitability of particle swarm optimisation for training neural hardware

Abstract

Learning algorithms implemented for neural networks have generally being conceived for networks implemented in software. However algorithms which have been developed for software implementations typically require far greater accuracy for efficiently training the networks than can be easily implemented in hardware neural networks. Although some learning algorithms designed for software implementation can be successfully implemented in hardware it has become apparent that in hardware these algorithms are generally ill suited, failing to converge well (or at all). Particle Swarm Optimisation (PSO) is known to have a number of features that make it well suited to the training of neural hardware. In this paper the suitability of PSO to train limited precision neural hardware is investigated. Results show that the performance achieved with this algorithm does not degrade until the accuracy of the networks is reduced to a very small number of bits.