Infinity Norm Based Neural Network Algorithm for Principal Component Analysis

In this paper a simple infinity norm based neural network algorithm for estimation of the principal component is developed. It seems to be especially useful in applications with changing environment, where the learning process has to be repeated in on-line manner. Theoretical analysis shows the weight vector converges to the principal eigenvector asymptotically. In comparison with the existing algorithms, numerical simulation shows that the proposed algorithm demonstrates fast convergence and robustness for a slightly noisy Gaussian samples with some points having large magnitude and angle with respect to the principal direction.