UNIVERSAL APPROXIMATORS FOR PATTERN-CLASS IFICATION BASED ON THE EXAMPLE OF CHARACTER-RECOGNITION

Classification problems, which can be described by sample data, are of ten solved by statistical concepts. Based on decision theory, it can b e shown that the - in this case optimal - bayes classifier estimates t he aposteriori probabilities of the pattern generating process. Many c lassifiers - such as polynomial classifier, multilayer peceptron and r adial-basis functions - are universal approximators, i.e. depending on the given degrees of freedom they approximate arbitrarily well the ap osteriori probabilities. Besides this theoretical treatment, the above mentioned classifiers are compared on the practical example of handwr itten digit recognition, and the characteristics of the different appr oaches are pointed out.