Parameterized normalization: application to wavelet networks

We present a method to handle the standard data normalization process, which is often used in neural network applications. Normalization is studied in the case of wavelet networks, and we derive a dynamic interpretation of its influence, which can be extended to several neural models. We show that data normalization may be simulated and parameterized to avoid data preprocessing, so that the normalization process becomes either tunable or dynamically adaptable. Both cases are illustrated with benchmark applications. The main benefit of the proposed methods is a big reduction of the time of convergence on a satisfactory classification rate