Neural metrics - Software metrics in artificial neural networks

Backpropagation based supervised feedforward Artificial Neural Networks (ANNs) have been developed for many applications (e.g. Rumelhart et al. 1986, Hinton 1989 Werbos 1990 and Riedmiller 1994) but no derailed study of the measurement of the qualify characteristics (e.g. complexity and efficiency) of the network system has been made. Without an appropriate measurement, it is difficult to tell how the network performs on given applications. In addition it is difficult to provide a measure of the algorithmic complexity of any given application This paper proposes a new set of software metrics, named Neural Metrics, which provide indicative measures of the quality characteristics of ANNs. Neural metrics that are non-primitive in nature are defined mathematically as Neural Metrics Functions.