Skip to Main Content
This paper provides a summary of a feasibility study conducted to assess and compare a weight based and a network output sensitivity based saliency measure for use with an Elman recurrent neural network (RNN). An experiment was designed to assign temporal data with significant noise, autocorrelation and crosscorrelation into one of two classes. To improve classification accuracy, feature saliency screening was performed to select a subset of the eight candidate input features using a weight based signal-to-noise ratio and an output sensitivity based measure. With consistent selection and ranking of features observed between the two saliency measures, both indicated a parsimonious subset of three of the original eight input features should be retained. Using CPU time as a surrogate measure of operations required, the computational efficiency was also found equivalent, with an observed difference of less than 2.5% between methods. Numerical results show a parsimonious subset of features improved generalization by significantly reducing the classification accuracy variance for multiple data sets and trained RNNs across time periods. An increase in classification accuracy for the last time period was even obtained for an independent validation set using the reduced feature set.