Performance Comparison of Neural Network Parallelization Techniques and Its Application for Large Data Classification

The parallelization of artificial neural network (ANN) classification for large data is still an open issue where the capability ANN is limited by the computing time and complexity rather than sample size. Training associated error usually get slower in order to find learning gradient direction that lead to time consumption and higher cost. Most of ANN parallelization focuses on application of physical or hardware strategy rather than the improvement of algorithm in order to optimize the classification. This paper proposes the enhancement of parallel ANN algorithm in order to obtain improved the performance classification and scalability. The classification task is enforced by individual cluster-based using partitioning method which represents a cluster of ANN classifiers. The classifier is developed based on back propagation ANN which adapts and test several training protocol of three different algorithms as a whole that gives variety of performances. Result shows significant improved of performance for speedup to more than 95% with complexity of O(n^k ). The results also proved that the proposed technique of scalability is efficient for ANN generalization ability for large dataset.