动态增殖流形学习算法.pdfVIP

计算机研究与发展 ISSNCN11．1777|TP JournalofComputerResearchandDevelopment 44(9)：1462～1468，2007 动态增殖流形学习算法 曾宪华 罗四维 。(北京交通大学计算机与信息技术学院 北京 100044) (西华师范大学计算机学院 南充 637002) (xianhuazeng2005@163．com) A DynamicallyIncrementalM anifoldLearningAlgorithm ZengXianhua ·andLuoSiwei 。(SchoolofComputerandInformationTechnology，BeljingJiaotongUniversity ， Beijig100044) (SchoolofComputerScience，ChinaWestNormalUniversity，Nach0g637002) Abstract Themaingoalofmanifoldlearningistofindasmooth low—dimensionalmanifoldembedded in high—dimensionaldataspace．Atpresent，manifoldlearninghasbecomeahotissueinthefieldofmachine learninganddatamining．In ordertoseek valuableinformationfrom high—dimensionaldatastream and largescale data set， it isurgently necessary to incrementally find intrinsic low—dimensionalmanifold structureinsuchobservationdataset．But，currentmanifoldlearningalgorithmshavenoincrementalability andalsocannotprocessthegiantdataseteffectively．Aimingattheseproblems，theconceptofincremental manifoldlearningisfirstlydefinedsystematicallyinthispaper．Itisadvantageoustointerpretthedynamic processofdevelopingastableperceptionmanifoldandtoguidetheresearchofmanifoldlearningalgorithms which fit to incrementallearning mechanism in man brain．According to the guiding principles of incrementalmanifold learning，adynamically incrementalmanifold learningalgorithm isthen proposed， which can effectively processthe increasing data setsand the giantdata setsampled from the same manifold．The novelmethod can find the global low——dimensionalmanifold by integrating the low．． dimensionalcoordinatesofdifferentneighborhoodobservationdatasets．Finally．theexperimentalresults onbothsynthetic “Swiss—roll”datasetandrealfacedatasetshow thatthealgorithm isfeasible． Keywords manifold learning；perception ma