Abstract
In any data set there some of the data will be bad or noisy. This study identifies two types of noise and investigates the effect of each in the training data of backpropagation neural networks. It also compares the mean square error function with a more robust alternative advocated by Huber.