ABSTRACT
In the Chapters 2, 3 and 4, we have studied the convergence of PCA learning algorithms. However, to apply PCA neural networks to practical problems, an important issue to address is how to determine an appropriate number of principal directions for the PCA neural networks. The number of principal directions crucially affects the computation results of PCA neural networks. For example, in the application of PCA neural networks to image compression, if the number of principal directions is too low, the quality of the reconstructed image would be inadequate. However, if the number of principal directions is too high, an excessive computational burden would be incurred. So far, this problem, that is, how to adaptively select an appropriate number of the principal directions so as to strike an adequate balance between the quality and complexity of computations, is still largely unsolved.
