Comparison of kernel estimators, perceptrons and radial-basis functions for OCR and speech classification