Appropriate Selection for Numbers of neurons and layers in a Neural Network Architecture: A Brief Analysis

Abstract

Identification of optimal number of neurons and layers in a proposed neural architecture is very complex for the better results. The determination of the hidden layer number is also very difficult task for the proposed network. The recognition of the effective neural network model in terms of accuracy and precision in results as well as in terms of computational resources is very crucial in the community of the computer scientists. An effective proposed neural network architecture must comprise the appropriate numbers of perceptrons and number of layers. Another research gap was also reported by the researchers community that the perceptron stuck during the training phase in finding minima or maxima for stochastic gradient to solve any engineering application. Therefore to resolve the problem of selection of neurons and layers an analysis was performed to evaluate the performance of the neural network architecture with different neurons and layers on the same data set. The results revealed that the justified network architecture would contain justified number of neurons and layers as more number of neurons and layers increase more computational resources and training time. It was suggested that a neural network architecture should be proposed comprising of minimum 2 to 5 layers. Entropy and Mean square error was considered as a yardstick to measure the neural network architecture performance. Results depicted that the an effective neural network architecture must initially be simulated or checked with minimum number of instances to evaluate the model.