Bernoulli error measure approach to train feedforward artificial neural networks for classification problems

The training of artificial neural networks usually requires that users define an error measure in order to adapt the network weights to achieve certain performance criteria. This error measure is very important and sometimes essential for achieving satisfactory solutions. Different error measures have been used to train feedforward artificial neural networks, with the mean-square error measure (and its modifications) being the most popular one. In this paper, the authors show that the Bernoulli error measure is very suitable for training feedforward artificial neural networks to learn classification problems. The authors compare the Bernoulli error measure with the popular mean-square error measure in terms of error surfaces, adaptation rates, and stability regions. The AND and XOR classification problems are used to illustrate the differences between the Bernoulli error measure and the mean-square error measure