Investigation of generalization ability by using noise to enhance MLP performance

The multilayer perceptron (MLP) is successfully used in many nonlinear signal processing applications. The backpropagation learning algorithm is very useful for various problems. But the MLP obtains low generalization ability if the number of hidden units is very large in training. In this paper, the authors show that if the MLP is trained with adding noise to hidden units, it obtains good generalization ability for any number of hidden units.