To train a neural network you need some measure of error between computed outputs and the desired target outputs of the training data. The most common measure of error is called mean squared error. However, there are some research results that suggest using a different measure, called cross entropy error, is sometimes preferable to using mean squared error.