A neural approach to data compression and classification

Abstract

Recently, neural networks have evolved as an alternate approach instead of rule-based systems for data compression and automated solution of interpolation or classification problems. The most prominent feature of the neural processing paradigm is its inherent adaptability permitting fairly easy modification of a neural system to perform in a wide range of application environments. This paper presents the cosine classifier, a neural network model designed for unsupervised adaptation and solution of classification problems. Classification of hand-written digits is used to demonstrate its performance.