Parallel learning of feedforward neural networks without error backpropagation

Abstract

A parallel architecture of the steepest descent algorithm for training fully connected feedforward neural networks is presented. This solution is based on a new idea of learning neural networks without error backpropagation. The proposed solution is based on completely new parallel structures to effectively reduce high computational load of this algorithm. Detailed parallel 2D and 3D neural network learning structures are explicitely discussed.