A projection filter for use with parameterised learning models

This paper presents a projection filter for use with parameterised learning models. Two aspects of the projection filter are considered. Firstly, the filter's operation is demonstrated using the perceptron learning rule on a simple two class discrimination problem. Secondly, the projection filter is used to extend the learning capabilities of a nonlinear spatio-temporal neural network model. An experiment was undertaken to compare the effectiveness of applying temporal backpropagation to a multilayer feedforward network employing either finite impulse response (FIR) or projection filters. Results show that the projection filter reached a lower mean square error (MSE) when compared to the FIR version of the network.