Fast backpropagation for supervised learning

In this paper, fast backpropagation (Fbp), a new, simple and computationally efficient variant of the standard backpropagation, is proposed. It continuously adapts the learning rate parameter /spl epsiv/, for individual synapses, using only network variables, without any significant increase in circuit complexity. The method is related to Fermi-Dirac distribution which is based upon quantum principles. The 'mean' update procedure employed offers a fascinating degree of stability and robustness. Even on individual runs Fbp, on average, converges quicker, particularly for non-Boolean inputs, and generalizes better than Quickprop with an identical set of initial random weights.