1] E.m. Kleinberg, an Overtraining-resistant Stochastic Modeling Method for Pattern Recognition

17 a similar reason, we did not nd the method of projection pursuit 5] in exploratory data analysis useful in this context. 4 Conclusions We described a method for inferring from the training data faithful but concise representations of the empirical class-conditional distributions. In doing this, we have abandoned many usual simplifying assumptions about the distributions: e.g. that they are simply-connected, unimodal, convex, or parametric (e.g. Gaussian). We have shown that a classiier can be constructed using a metric deened on these distribution maps. Our method requires unusually large and representative training sets, which we provided through pseudo-random generation of training samples using a realistic model of printing and imaging distortions. We illustrated the method on a challenging recognition problem: 3755 character classes of machine-print Chinese, in four typefaces, over a range of text sizes: in a test on over three million images, the perfect-metric classiier achieved better than 99% top-choice accuracy. In addition, we showed that it is superior to a conventional parametric classiier using comparable resources. The reject behavior, permitted by the distinctive distributions of distances to correct or incorrect classes, is also very reliable. We have also shown a way to construct similar feature transformations and classiiers for arbitrary domains. The features and the metric were derived with minimum heuristics. Moreover, the classiier's accuracy can be improved with additional training data. In this study we have concentrated on linear mappings as features. Future studies may focus on investigating other families of features, especially those that can be used under weaker conditions on the class-conditional distributions. Acknowledgements We would like to thank Eugene Kleinberg for the communications of his theory, and Pavlidis for their helpful comments. The font descriptions used in the experiments were supplied by William Sun along with his software packagèGB2PS' distributed on Internet. We are thankful for his contribution.