A fast learning algorithm for principal component extraction with data dependent learning rate
Lijun Liu, Rendong Ge, Jun Tie · 2010
We propose a fast adaptive learning algorithm for computing principal eigenvector of covariance matrix arisen in the field of signal processing, where the learning process has to be repeated in online manner. Compared with most existing neural algorithms, the proposed approach effectively makes use of the online estimation of eigenvalue to update the principal eigenvector, which makes the method works with an adaptive data dependent learning rate and thus demonstrates a fast convergence speed. Numerical experiment further shows that this data dependent learning rate in the proposed algorithm offers significant advantages over that of constant learning algorithm.