Autonomous learning algorithm for fully connected recurrent networks

In this paper, fully connected RTRL neural networks are studied. In order to learn dynamical behaviours of continuous time processes or to predict numerical time series, an autonomous learning algorithm has been developed. The originality of this method consists in the gradient-based adaptation of the learning rate and time parameter of neurons using a small perturbations method. Starting from zero initial conditions (neural states, rate of learning, time parameter and matrix of weights) the evolution is completely driven by the dynamic of the learning data. Stability issues are discussed, and several examples are investigated in order to compare the performances of the adaptive learning rate and time parameter algorithm with the constant parameters one.