New Approach of the Recurrent Neural Network Training

Abstract

In this work the technique o f creation o f adapthre training algorithms for recurrent neural networks (RNN) is cortsidered. These algorithms have high convergence and accuracy on a comparison with traditional backpropagation. The original technique of calculation of an adaptive training step with use of steepest descent method is resulted. The features of calculation of an adapttve pitch for neural elements with recurrent connections are discussed. Are considered the neural units with various functions of activation, used in architectures neural systems of forecasting. The indicated computing experiments demonstrate advantage o f the developed RNN training methods.