A Competitive Wavelet Layer for Pattern Clustering

A competitive “wavelet layer” is proposed for pattern clustering. It exploits the representation capabilities of adaptive wavelets to generate template approximations for each cluster of data. A brief review of adaptive wavelet representations, as well as some insight into local minima problems, is provided. The method is illustrated by a simple clustering problem, in which step responses of dynamic systems are discriminated with basis on the presence of parasitic oscillations. The results suggest that the wavelet layer exhibits superior performance than the conventional competitive neural layers when patterns exhibit a low signal-to-noise ratio.