Error splitting method for training a built-in neural network storage element with SR-latch functionality

Abstract

A method for dividing the error when passing through a storage element with SR-latch functionality for embedding in neural networks with one output and two inputs is presented. In the backpropagation method, the error is divided from one output to the two inputs with opposite coefficients. In this way, the element trains the previous neural layers to detect input vectors, which indicate the beginning and the end of statuses to be remembered. The forward and reverse interfaces of the element are defined, respectively for calculating the information signal and the error.