On-line training of a continually adapting adaline-like network

Abstract

A rigorous mathematical analysis is presented of an adaline-like network operating in continuous time with spatially continuous inputs and outputs. Weights adapt continually, whether or not a training signal is present. It is shown that consistent input-output pairs can be learned perfectly provided every pattern is repeated at least once in every N successive inputs, and the input patterns are nearly orthogonal, depending on N.