Cascading classifiers

We propose a multistage recognition method built as a cascade of a linear p arametric model and a k-nearest neighbor (k-NN) nonparametric classifier. T he linear model learns a "rule" and the k-NN learns the "exceptions" reject ed by the "rule." Because the rule-learner handles a large percentage of th e examples using a simple and general rule, only a small subset of the trai ning set is stored as exceptions during training. Similarly during testing, most patterns are handled by the rule-learner and few are handled by the e xception-learner thus causing only a small increase in memory and computati on. A multistage method like cascading is a better approach than a multiexp ert method like voting where all learners are used for all cases; the extra computation and memory for the second learner is unnecessary if we are suf ficiently certain that the first one's response is correct. We discuss how such a system can be trained using cross validation. This method is tested on the real-world application of handwritten digit recognition.