6.1 Introduction
In Chapter 5, I discussed several adaptive algorithms for computing principal and minor eigenvectors of the online correlation matrix Ak∈ℜnXn from a sequence of vectors {xk∈ℜn}. I derived these algorithms by applying the gradient descent on an objective function. However, it is well known [Baldi and Hornik 95, Chatterjee et al. Mar 98, Haykin 94] that principal component analysis (PCA) algorithms based on gradient descents are slow to converge. Furthermore, ...