Reliability analysis of artificial neural networks

Neural network technology has been applied to a variety of science and engineering problems that involve the extraction of useful information from complex uncertain data. The problem of estimating the reliability of a neural network is discussed. Reliability is defined as the probability that a correct output is produced by the neural network, even though some of the constituent components have failed. The methodology described is applicable to a large variety of neural networks, and can incorporate a number of alternative failure models. An example network is analyzed to show the effect of component failures and the number of training patterns on reliability. Simple methods to improve reliability are investigated