Enhancing the Performance of the BackPropagation for Deep Neural Network

The standard Backpropagation NeuralNetwork (BPNN) Algorithm is widely used insolving many real problems in world. But thebackpropagation suffers from different difficultiessuch as the slow convergence and convergence tolocal minima. Many modifications have beenproposed to improve the performance of thealgorithm such as careful selection of initial weightsand biases, learning rate, momentum, networktopology and activation function. This paper willillustrate a new additional version of theBackpropagation algorithm. In fact, the newmodification has been done on the error signalfunction by using deep neural networks with morethan one hidden layers. Experiments have beenmade to compare and evaluate the convergencebehavior of these training algorithms with twotraining problems: XOR, and the Iris plantclassification. The results showed that the proposedalgorithm has improved the classical Bp in terms ofits efficiency.