An on-line unsupervised learning machine for adaptive feature extraction

An on-line unsupervised learning machine (LEAP) for adaptive feature extraction (principal component decomposition of a stochastic process) is introduced. Some results on convergence analysis of the LEAP system are summarized. A result on the identification of a nontrivial domain of attraction for the system is given. A new application is shown. LEAP can be extended to perform on-line higher-order statistics component decomposition of a stochastic process. This extended system (ELEAP) can then be combined with two identification techniques resulting in an on-line signal processing system for parameter estimation of multichannel moving-average processes using higher-order statistics.<>