Hybrid neural network for efficient training

Abstract

Neural Network is a platform to implement artificial intelligence and universally used neural network is Backpropagation (BP). Local minimum, slower convergence, premature saturation, training pattern overspecialization limits the performance of Backpropagation algorithm. To withstand those problems several modified algorithms was proposed. A faster superintendent algorithm named Hybrid Backpropagation algorithm (HBP) is used in this paper because of training of the neural network is found from Back-propagation with Chaotic Learning (BPCL) with different types of chaos, Maximization of gradient function (BPfast) and Error back-propagation (EBP) to mitigate the limitations of BP. HBP is examined on several benchmark categorize problems like glass, breast cancer, diabetes, horse and Australian credit card. For the ability of generalization and fast rate of convergence than any single algorithm HBP outperforms BP, BPfast, BPCL and EBP.