Training Arbitrarily Connected Neural Networks with Second Order Algorithms

Abstract

Neural networks have been an active area of research and application for many years. Today they are gaining popularity with the growing processing power of modern computers. With neural networks gaining popularity comes a demand for a simple and reliable method of training all types of networks which is the focus of this paper. Neural Network Trainer is a training package that allows the user to create a simple netlist style network architecture in a text file and quickly begin training. Several algorithms are implemented including Error Back Propagation as well modified versions of the Levenberg-Marquardt algorithm. The software is demonstrated with results verifying the implemented algorithms as well as the trained neural networks.