Study on Optimization of Artificial Neural Network Generalization Power Based on Architecture
Artificial Neural Networks (ANNs) have attracted considerable attention from researchers in many fields, including economics, medicine, data processing, robotics, quality control, optimization and security, and have been used to solve a wide range of problems. The discovery of numerous applications and promising benefits to a diverse field encourages studies in this area. Architecture of the network is one of the major challenge in the field of Artificial Neural Network. Hidden layers plays a vital role in the performance of ANN causing the problem of either underfitting or overfitting under certain circumstances. Overfitting is the phenomenon wherein the network fits the training data set so well that even noise of the training data are memorized causing performance to drop when it is tested against unknown data set. Underfitting, the opposite of over-fitting, occurs when the model is incapable of capturing the variability of the data and the resulting model will have suboptimal predicative ability. The structure of the network also has direct effect on training time and classification accuracy. Although there are some discussions in the literature of the impact of network structure on the performance of the ANN, there is no standard method or approach for selecting the optimum structure. In this paper, the relationship between the number of hidden layers, nodes (neurons) and the accuracy of the classification is investigated. The result of this study can be used as a guideline for the selection of design parameter in setting artificial neural networks.
Keywords- Artificial Neural Network, Network Architecture, Hidden Layer, Neurons.