Two Classes of Non-Parametric Techniques for Pattern Recognition and Their Error Analysis

摘要

Another criterion is added to handle the case when distributional information is lacking. This criterion is to approximate the Bayes Solution based only on the statistics acquired during the learning phase. This criterion approaches the optimum risk decision as the learning phase is increased. Two classes of non-parametric techniques are proposed. Corresponding error analyses for these two techniques are made in order to determine how much is lost by using sub-optimal (i.e., a finite learning phase) decisions.