IMPROVEMENTS TO THE BACKPROPAGATION ALGORITHM

Abstract

This paper presents some simple techniques to improve the backpropagation algorithm. Since learning in neural networks is an NP-complete problem and since traditional gradient descent methods are rather slow, many alternatives have been tried in order to accelerate convergence. Some of the proposed methods are mutually compatible and a combination of them normally works better than each method alone.

Keywords

• Artificial Neural Network (ANN)
• backpropagation
• extended network