Fast convergence of the backpropagation learning algorithm by using adaptive accuracy of weights

The authors investigate increase in the speed of convergence for the standard backpropagation learning algorithm by adapting the accuracy used to represent the weights during the learning process, instead of using a fixed weight accuracy. The change of the accuracy is determined adaptively from the learning process itself, using statistical measures of the error function. Simulations were performed and results are presented for a XOR problem, relating to seven-segment liquid crystal diode coding and a 8-5-8 encoder. Simulation results show that convergence speed can be increased drastically