Neural network performance versus network architecture for a quick stop training application

The relationship between neural network (NN) performance and the parameters that form the architecture of the neural network is complicated. This paper studies an issue that is still controversial—the effect of the number of neurons and number of hidden layers on NN performance. This paper tests if the above parameters affect network performance when the NN is trained according to the quick stop strategy. It was determined that the number of neurons did not have any effect on the NN performance for a single layer or a double layer network. It was also found that the training error and calibration error could be poor indicators of the prediction correlation coefficient.