Validation Based Cascade-Correlation Training of Artificial Neural Networks

Abstract: The Cascade-Correlation learning algorithm, Cascor, is a self-constructive learning algorithm introduced by Fahlman in 1990, which builds the networks one node at the time based on the errors of the network. Cascor has been criticized for creating excessively deep networks and easily overfit. This thesis investigates a few methods that use an internal validation set to add nodes that perform better on unseen data. The goal is to improve the generalization of the networks and examine how these methods affect the network architecture. It is shown that the investigated methods are all able to reduce the depths of the networks and decrease the overfitting of large networks.