An Improvement Of Classifers’ Leanring Ability And Generalization In Pattern Recognition

Pattern recognition is a procedure of processing and analyzing various forms ofinformation that characterize the object or phenomenon, to make the description,identification, classification and interpretation for things or phenomena. Patternrecognition technologies can be used for face recognition, fingerprint recognition,speech recognition and document classification. Building classifiers according totraining samples is an important part of statistical pattern recognition. It is found thatthere is a contradiction between learning ability and generalization when building aclassifier. In the theory of statistical learning, the VC dimension proposed by Vapnikbased on hypotheses set gives the relationship between the generalization and learningability of the learner. It can be strictly considered that a smaller VC dimension of anassumption collection leads to stronger generalization and weaker learning ability. Thusthe contradiction when designing a classifier is that it is difficult to simultaneouslyimprove learning ability and generalization. So balance and compromise is necessary inthe practical work.The new classification method proposed in this paper is designed to reduce suchcontradiction. Building a classifier can be described as a procedure of searching aassumption, which make the best classification of training data, from the assumptioncollection. The analysis of generalization in statistical theories is based on the bias(VCdimension) of assumption collection, and the description of learning ability is also basedon the bias. This is one reason for the contradiction. A method for reducing thecontradiction is changing generalization and learning ability according to differentfactors. The basic idea of this paper is that learning ability is based on the assumptioncollection and generalization capability is based on the searching strategy. Unbiasedassumptions enhance the learning ability of classifiers, and a stronger searching strategyenhances the generalization ability of classifiers.The thesis of this paper includes the following four parts:①A searching bias named classification surface with maximum margin isproposed. A special class of data distributions called pure distribution pair and dualpoint pair based on it is defined.(Different from supporting vector machine, theclassification surface with maximum margin is surface related to the distribution ratherthan flat.) ②A method is proposed to find approximate dual point pair in the trainingsamples. The strategy based on these dual point pairs is designed for establishingclassification surface and classifying.③A method is proposed to make data distribution meet the pure distribution pairby eliminating noise and transforming the data.④The classification method is implemented. The experimental results show thatthis method is better than SVM to some extent for solving the problem of the highlynonlinear data distribution.