Canonical form of Recurrent Neural Network Architecture
N. Selvanathan, Mashkuri Hj. Yaacob · Industrial and Engineering Applications of Artificial Intelligence and Expert Systems · 2022
A recurrent neural network (RCNN) allows feedback from the output of a neuron to its own input, input to neurons of the same layer or neurons of other layers. Feedforward network allows output of neurons in a particular layer to be fed to inputs in the forward layers, hence the excitation is propagated forward. Recursive network allows feedforward and feedback and its recursive nature allows sustained activity, without any external influence. RCNN are able to store information and are suitable for time series prediction. In addition, it is applied for speech processing especially for sequence to label or vice versa application. The mathematical analysis of RCNN is complex in virtue of its feedforward and feedback nature and various researchers have made a attempts to model its dynamics [ NC83 ]. This paper is based on the Nerrand et al model [ A94 ] and is the most general type of recurrent neural network that has been proposed. An attempt is made to obtain the canonical representation of the Nerrand et al architecture and its dynamics written in the state space formulation.