Hyper surface perceptron and its applications

This paper deals with the problems associated with the perceptron model used in neural networks. The concept of preprocessing input signals is used before the linear summation via perceptron weights. The input signals of a specific pattern of length m are augmented into corresponding pattern of length 2^m. The augmented patterns are orthogonal and this simplifies greatly the perceptron learning and solves the inseparability problem of the input signal patterns. The paper introduces a complete analysis for the mathematical model of the proposed perceptron and weight estimation process. The most important result is that a relatively simple iterative method is needed for weight estimation. Also, the weights are limited in the range (-1 to +1) and change in a quantised manner with resolution (1/2^m). The paper presents many practical applications to prove the applicability of the approach.