Performance evaluation and dynamic node generation criteria for “principal component analysis” neural networks

This paper is concerned with the solution of the principal component analysis (PCA) problem with the aid of neural networks (NNs). After an overview of the basic NN-based PCA concepts and a listing of the available algorithms, two criteria for evaluating PCA NN algorithms are proposed. Then, a new criterion for the generation of improved PCA NN structures with reduced size is presented. Using this criterion, one can start with a small network and dynamically add new nodes at the hidden layer(s) during training, one at a time, until the desired performance is achieved. A simulation example is provided that shows the applicability and effectiveness of the methodology.