Adaptive learning using higher-order statistics

The classification of random and deterministic signals is considered. A cumulant-based classifier is derived, which is insensitive to additive Gaussian noise and can classify non-minimum phase signals. A computationally efficient implementation structure is proposed, which avoids the explicit computation of cumulants. Its performance analysis is discussed. Adaptive training algorithms for the classifier are also derived, using gradient methods to minimize cumulant-based criteria. Gaussian noise insensitivity is preserved. The authors prove global convergence, irrespective of the initial conditions, and illustrate the tracking capabilities with simulations.