Sequential Blind Signal Extraction
Andrzej S Cichocki, Шун-ичи Амари · 2002
There are three main objectives of this chapter: (a) To present simple neural networks (processing units) and propose unconstrained extraction and deflation criteria that do not require either a priori knowledge of source signals or the whitening of mixed signals. These criteria lead to simple, efficient, purely local and biologically plausible learning rules (e.g., Hebbian/anti-Hebbian type learning algorithms). (b) To prove that the proposed criteria have no spurious equilibriums. In other words, the most learning rules discussed in this chapter always reach desired solutions, regardless of initial conditions (see appendixes for proof). (c) To demonstrate with computer simulations the validity and high performance for practical use of the derived learning algorithms. In this chapter two different models and approaches are used. The first approach is based on higher order statistics (HOS), which assume that sources are mutually statistically independent and they are non-Gaussian (expect at most one) and as criteria of independence, we will use some measures of non-Gaussianity. The second approach based on the second order statistics (SOS) assumes that source signals have some temporal structure, i.e., the sources are colored with different autocorrelation functions or equivalently different shape spectra. Special emphasis is given to blind source extraction (BSE) in the case when sensor signals are corrupted by additive noise using the bank of bandpass filters.