Curvature-driven smoothing in feedforward networks

Summary form only given. The standard backpropagation learning algorithm for feedforward networks aims to minimize the mean square error defined over a set of training data. This form of error measure can lead to the problem of over-fitting in which the network stores individual data points from the training set, but fails to generalize satisfactorily for new data points. In the present work, the author proposes a modified error measure which can reduce the tendency to over-fit and whose properties can be controlled by a single scalar parameter. The proposed error measure depends both on the function generated by the network and on its derivatives. A novel learning algorithm was derived which can be used to minimize such error measures