Evidence of Efficiency of Recurrent Neural Networks with ARMA-like Units

摘要

We study in this paper a recurrent neural model where we associate an ARMA process to each neuron-like unit. In order to quantify the effects of the new free parameters on the network, we will use a statistical tool, the analysis of variance (or ANOVA) to establish our results. Finally, practical results highlighting the improvements are presented through the prediction of the Mackey-Glass chaotic signal.