Recognition of class membership by means of weak, statistically dependent features /

A method of automatic classification is developed for the case in which the features used to determine the class of an unknown object x are individually weak. The features are weak in the sense that any subset of the universe defined by a single feature value of x contains many objects belonging to a class different from that of x. The classes are defined by a small collection of examples, which are objects whose class membership and feature values are known. The basic problem in classification by example is the estimation of probabilities from a small number of examples. Ideally, the class of x should be determined by estimating the class probabilities in the subset J defined by the conjunction of all of the feature values of x. But J usually will contain no examples on which to base an estimate of the class probabilities. The recognition procedure offered here employs a subset D defined by the conjunction of some, but not all, of the feature values of x. It is postulated that the features are statistically dependent in such a way that the particular kind of dependence exhibited by the features defining D can be used to infer the class of x. This recognition procedure includes a hypothesis-testing procedure in which the competing hypotheses are related to the dependence of the features defining D and are only indirectly related to the classes to which x could possibly belong. In this approach the class of x can be decided without requiring that D contain any examples. These principles are demonstrated by experiments in character recognition.Includes bibliographic references (page 39).A method of automatic classification is developed for the case in which the features used to determine the class of an unknown object x are individually weak. The features are weak in the sense that any subset of the universe defined by a single feature value of x contains many objects belonging to a class different from that of x. The classes are defined by a small collection of examples, which are objects whose class membership and feature values are known. The basic problem in classification by example is the estimation of probabilities from a small number of examples. Ideally, the class of x should be determined by estimating the class probabilities in the subset J defined by the conjunction of all of the feature values of x. But J usually will contain no examples on which to base an estimate of the class probabilities. The recognition procedure offered here employs a subset D defined by the conjunction of some, but not all, of the feature values of x. It is postulated that the features are statistically dependent in such a way that the particular kind of dependence exhibited by the features defining D can be used to infer the class of x. This recognition procedure includes a hypothesis-testing procedure in which the competing hypotheses are related to the dependence of the features defining D and are only indirectly related to the classes to which x could possibly belong. In this approach the class of x can be decided without requiring that D contain any examples. These principles are demonstrated by experiments in character recognition.Research supported by Aerospace Medical Division, Air Force Systems Command, United States Air Force; report prepared by Radio Corporation of America, RCA Laboratories, under contract no.Mode of access: Internet