Perceptron based neural network predictors in lossless data compression

This paper proposes the use of perceptron based neural networks in a new application for lossless data compression in the two-stage predictor-encoder scheme. Different single- and multi-layer perceptrons are introduced into the first stage. Architectural, neurodynamic, training, transmission and optimisation issues involving the perceptron implementations are discussed. The networks are tested using a variety of telemetry data files. Arithmetic coding is employed in the second stage and the two-stage schemes are evaluated. Existing schemes using the FIR, NLMS and RLSL predictors are introduced as comparison benchmarks. It is found that the performance of the perceptron schemes are comparable and even better than the existing methods using the linear predictors. It is also found that in certain cases, single-layer neural networks are capable of performing better-than multi-layer networks, and that performance trends vary when the predictors are applied in the single-stage and two-stage schemes.