Model based classification of transient signals using the MLANS neural network

Abstract

A maximum likelihood artificial neural system (MLANS) neural network is proposed for transient signal recognition. The MLANS learning efficiency greatly exceeds that of other neural networks and is approaching the information-theoretical limit on performance of any neural network or algorithm. The MLANS operates on a two-dimensional representation of the signal in either the short-term spectral or the Wigner transform domain. The first layer of the network uses structured second-order neurons to estimate the signal model from training data. A second layer performs optimal multimodal Bayes classification. Learning efficiency approaching the information-theoretical limit is achieved in each layer of the MLANS.<>