A solution of speed-stability problem in noncoupled neural network algorithms

A major and common problem, which is the so-called speed-stability problem, exists in most noncoupled neural network based algorithms for principal (PCA) and minor (MCA) component analysis. Coupled PCA or MCA algorithms, in which the principal or minor eigenvector and the corresponding eigenvalue of a covariance matrix are estimated in coupled equations simultaneously, can mitigate the speed-stability problem. Or in other words, coupled algorithms perform better, i.e., converge faster and more stable, than noncoupled algorithms. Until now, a large number of noncoupled learning algorithms have been proposed and analyzed, but unfortunately, there are only few of coupled learning algorithms can be found in existing literatures. To address issue, in this paper, we propose a method of solving the speed-stability problem to improve the convergence speed and stability of noncoupled algorithms, which is actually done by converting the noncoupled algorithms to the coupled ones. Based on the proposed method, we can obtain lots of coupled algorithms from the existing noncoupled algorithms, and all of them do not have the speed-stability problem thus perform better than their noncoupled versions.