A convergent neural network learning algorithm

A globally guided backpropagation (GGBP) training algorithm is presented. This algorithm is a modification of the standard backpropagation algorithm. Instead of changing a weight w_ij according to the partial derivative or error, E, with respect to w_ij, an attempt is made to minimize E in the output space. The change in weights W is computed based on the desired changes in the output O. The new algorithm is an analog to backpropagation with a dynamically adjusted learning rate η. This learning rate changing scheme avoids the problems associated with a heuristic learning rate adjusting method. Two main advantages of GGBP are fast learning speed and convergence to a global optimal solution