Robust low-sensitivity Adaline neuron based on Continuous Valued Number System

Abstract

In this article Continuous Valued Number System is studied as an alternative method for implementing Analog Neural Networks. Continuous Valued Number System is analog in nature and employs digit level analog modular arithmetic. The information redundancy among the digits allows efficient operations using analog circuitry with arbitrary accuracy. The general operations in this number system are more precise than regular analog operations, thus enabling us to implement large size analog neural networks with more precision. In this article, function evaluation properties of the Continuous Valued Number System are introduced. These key properties are used for developing analog Adaline with a nonlinear activation function. Stochastic modeling of a network of such elements is carried out which indicates that the proposed network has low sensitivity to implementation errors.