一类改进的最小距离分类器的增量学习算法-模式识别与人工智能.pdfVIP

第20 卷摇 第 3 期摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 模式识别与人工智能摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 Vol.20摇 No.3 摇 2007年6月摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 PR AI摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 Jun摇 摇 2007 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 一一类改进的最小距离分类器的增量学习算法* 桑摇 农摇 摇 张摇 荣摇 摇 张天序 (华中科技大学 图像识别与人工智能研究所 武汉摇 430074) 摘摇 要摇 提出一种基于改进的最小距离分类器的增量学习算法,消除增量学习过程中产生的分类器内部结构的相 互干扰,使分类器既能记住已学习的知识,又能学习新知识.增量学习需要对分类器结构进行调整,必须使用有代表 性的已学习样本帮助分类器在学习新知识时复习旧知识.针对正态分布的样本集提出一种筛选算法,只保留有代表 性的少量样本,大大减少存储消耗和重新训练的计算开销.实验结果证明该算法对样本的识别准确率高,在有效识 别新样本的同时对以前学习的样本也保持较高的识别率,消耗存储空间小. 关键词摇 增量学习,最小距离分类器,干涉灾难,样本筛选 中图法分类号摇 TP391.4 Incremental Learning Algorithm of a Modified Minimum蛳Distance Classifier SANG Nong,ZHANG Rong,ZHANGTian蛳Xu (Institutefor Pattern Recognition andArtificial Intelligence,Huazhong University of Science and Technology,Wuhan430074) ABSTRACT An incremental learning algorithm of a modified minimum蛳distance classifier isproposed to eliminate the inter disturbance of the classifier during the incremental learning process. It enables the classifier to remember the old knowledge and learn the new one at the same time. Incremental learning requires the modification of the classifier structure,and certain number of old samples must be reserved to help to review old knowledge while learning the new one. In terms of normal distributed sample sets, a new filtering algorithm isproposed. A few samples are preservedwhich are representative and greatly reduced the cost of storing as well as retraining. Experimental results show that the algorithm gives high rate of recognition correctness. Itmakestheoldsamplesremainhighrateofrecognitio