Radial basis function network with short term memory for temporal sequence processing

This paper discusses radial basis function networks extended with short term memory model. The short term memory model which is an essential component in temporal pattern recognition problems captures necessary features in time. We attempt to analyze the performances of a number of short term memory models and to study the sensitivity on the selection of their parameters. Since it is required a fairly good knowledge of the structure of a system to extract good features, we propose an architecture of a radial basis function network extended with a short term memory model whose outputs are fully connected to the input units of the radial basis function network. The proposed architecture makes it possible to decide which features in time should be used for a given task. We also develop a method to train the connection strengths from the memory model to the input units of the radial basis function network. Through simulation results, the proposed architecture is shown to be effective when we only have partial information about the structure of the system in advance