Multiresolution neural networks for recursive signal decomposition
Kai-Chiu Kan, Kwok‐Wo Wong · 2002
This paper proposes a novel algorithm for the synthesis of multiresolution neural networks that possesses self-construction capability. It is referred as the recursive variance suppression growth method. An explicit link between the network coefficients and wavelet transforms is found. By the proposed algorithm, the network is allowed to start with null hidden-layer neuron. As training progresses, the network grows autonomous to tackle the problem being studied. Simulations on a number of natural voice signals and a synthesized piecewise function show that clear and optimal local representation is obtained in the spatial-frequency spectrum. This indicates that the proposed approach is superior to the traditional signal decomposition techniques, especially for time-varying signal analysis.