Analysis of linear predictive data such as speech by a class of single-layer connectionist models

Summary form only given: A class of single-layer connectionist models is analysed where the nonlinearity includes the logistic function commonly used in error back propagation analysis. It is shown that when the input data can be modelled by a linear predictive or autoregressive process, a solution exists for the weights which minimises the cost function and hence an output error cost function. This establishes the weight sets for single-layer connectionist models whose input data are linear predictive processes, such as speech. Comparisons are made with the equivalent error back propagation analysis and an alternative network structure is described. The method also suggests a form of connectionist vector quantisation (CVQ) for the analysis of speech. A number of experimental results are given