Dynamic learning using exponential energy functions
M. Ahmad, F.M.A. Salam · 2003
The authors use a continuous-time gradient descent weight update law for supervised learning of feedforward artificial neural networks because of specific advantages over its discrete-time counterpart. An exponential energy function is used in the update law. It is shown that this energy function would speed up the learning dynamics and insure faster convergence to a useful minimum. The dynamics would 'skip' minima which are at higher energy levels and converge to one at a lower energy level. Software implementation of the learning dynamics based on the exponential energy function is described. Various supporting simulations on the XOR and character recognition problems are also included.>