Increment learning method on the basis of supporting vector geometrical significance

Abstract

The invention discloses an increment learning method on the basis of supporting vector geometrical significance. The method includes: providing a method for selecting given-number convex hull vortex samples, naming the method as CHVS, and theoretically proving the convex hull vortex samples selected by the method is the convex hull vortex samples; using CHVS in each category of sample sets, selecting the significant samples which can keep various categories of convex hull information to the maximum extent from large-data-amount training samples, and naming the method as VS; retraining the selected significant samples with new samples to obtain an updated categorizer. The method effectively realizing SVM increment learning and categorizing is universal. Compared with other classical SVM increment learning methods, the method is suitable for online increment learning tasks with large data amount and promising in application prospect.

Description

technical field [0001] The invention relates to the field of computer pattern recognition, in particular to an incremental learning method based on the geometric meaning of a support vector machine (SVM). Background technique [0002] Classifier design is the research focus in the field of computer pattern recognition, because classifiers are the basic tools of pattern recognition research. In general, the current classifier algorithm has two key problems: (1) the generalization ability of the classifier; (2) the time spent on training the classifier. [0003] The generalization ability of the classifier is the ability of the classifier to predict the unknown sample category, that is, the level of classification accuracy. Support vector machine, or SVM, is a machine learning method based on statistical learning theory. It is a concrete realization of the idea of ​​structural risk minimization and has good generalization ability. By introducing the kernel mapping method, it...

AI Technical Summary

Problems solved by technology

In general, the current classifier algorithm has two key issues: (1) the generalization ability of the classifier; (2) the time spent training the classifier

Although this type of method overcomes the defect that the previous type of method may lose important data samples during the incremental learning process and improves the classification accuracy, it requires many iterations to reach the stop condition when processing large-scale data, so it cannot meet the requirements of real-time online learning. requirements

Images