Learning generalization by validation set

Training a neural network to respond reasonably to input data not present in the training set is believed to be difficult. This is known as the generalization problem. We propose a learning method for solving this problem. The proposed method adjusts the number of hidden units in the network according to the difference between the errors for the validation and the training sets. The difference reflects the degree of generalization during the learning process. The correction of weights is based on the error back propagation. Numerical simulations of curve fitting demonstrate the effectiveness of our algorithm.