Learning of sinusoidal frequencies by nonlinear constrained Hebbian algorithms

The authors study certain unsupervised nonlinear Hebbian learning algorithms in the context of sinusoidal frequency estimation. If the nonlinearity is chosen suitably, these algorithm often perform better than linear Hebbian PCA subspace estimation algorithms in colored and impulsive noise. One of the algorithms seems to be able to separate the sinusoids from a noisy mixture input signal. The authors also derive another algorithm from a constrained maximization problem, which should be generally useful in extracting nonlinear features.<>