A Neural Architecture Based on the Adaptive Resonant Theory and Recurrent Neural Networks

In this paper, we propose a novel neural architecture that adaptively learns an input-output mapping using both supervised and non-supervised trainings. This neural architecture consists of a combination of an ART2 (Adaptive Resonance Theory) neural network and recurrent neural networks. For this end, we developed an Extended Kalman Filter (EKF) based training algorithm for the involved recurrent neural networks. The proposed ART2/EKF neural network is inspired in the visual cortex and the brain mechanisms. More precisely, the non-supervised ART2 neural network is used to coordinate specialized recurrent neural networks in a specific input space domain. Our aim is to design a neural system that learns in real time a new input pattern without retraining the neural network with the whole training set. The proposed neural architecture is used to adaptively predict the traffic volume in a computer network. We verify that the ART2/EKF is capable of finding patterns in the traffic time series as well as to obtain the transmission rate that should be made available in order to avoid byte losses in a computer network link.