CURVATURE-DRIVEN SMOOTHING - A LEARNING ALGORITHM FOR FEEDFORWARD NETWORKS

The performance of feedforward neural networks in real applications ca n often be improved significantly if use is made of a priori informati on. For interpolation problems this prior knowledge frequently include s smoothness requirements on the network mapping, and can be imposed b y the addition to the error function of suitable regularization terms. The new error function, however, now depends on the derivatives of th e network mapping, and so the standard backpropagation algorithm canno t be applied. In this letter, we derive a computationally efficient le arning algorithm, for a feedforward network of arbitrary topology, whi ch can be used to minimize such error functions. Networks having a sin gle hidden layer, for which the learning algorithm simplifies, are tre ated as a special case.