WSOM: Building Adaptive Wavelets with Self-organizing Maps

摘要

The WSOM (Wavelet Self-Organizing Map) model, a neural network for the creation of wavelet bases adapted to the distribution of input data, is introduced. The model provides an efficient on-line way to construct high-dimensional wavelet bases. Simulations of a lD function approximation problem illustrate how WSOM adapts to non-uniformly distributed input data, outperforming the discrete wavelet transform. A speaker-independent vowel recognition benchmark task demonstrates how the model constructs high-dimensional bases using low-dimensional wavelets.