Cost-performance evaluation of analog neural networks and high order networks

High order networks, studied over the past few years, have been shown to improve learning rates, increase storage capacity and reduce the number of layers required in comparison with first order nets. One issue which usually remains in the background, is the relative cost-performance of such nets. In this paper we address this issue in a more general framework, which we define, namely generalized high order networks. We present a cost-performance model and demonstrate its usability by analyzing some well-known first and high order networks. Our aim is to provide a simple, yet illuminating model, which enables the evaluation and analysis of generalized high order networks.