Supervised Learning with Neural Network Growth and Pruning Techniques

Project Outcomes

·       Development of new supervised learning approach which will allow the neural network to allocate a new hidden unit if the network performs poorly on a presented pattern in order to correct its response. It will be suitable for solving multi-classification problems and for on-line learning applications where the training data arrives sequentially in time as well as for off-line training. 

·       Development of a neural network topology which is able to control the complexity of its structure by allocating a new hidden unit to correct its response to the presented pattern. Fixed-size networks either use too few units or too many in which case the network demonstrates poor learning or poor generalization. The newly allocated units do not interfere with previously allocated units.

·       Extracting of a small number of supporting patterns from the training data that are relevant to the classification, which will further improve the convergence rate of the NN.

·       Comparison of results on multi-class classification problems with some standard methods for classification. Several benchmark data sets will be considered to show the classification performance on the training and test sets achieved by the proposed approach and some other neural network models.