面向机器学习的相对变换.pdf

ISSN 计算机研究与发展 1000．1239／CN11．1777／TP Researchand JournalofComputer Development 面向机器学习的相对变换 文贵华 (华南理工大学计算机科学与工程学院 广州 510641) (crghwen@scot．edu．cn) RelativeTransformationforMachine Learning WenGuihua China (SchoolComputerScienceand Universityof 510641) of Engineering，South Technology，Guangzhou Abstract machine suchasmanifold andthe vector Recentlydevelopedlearningapproaches learning support workwell thecleandatasetsevenifthesedatasetsare machine on learning highly aremuchsensitivetonoisesoroutlierscontainedinthedata thesenoisesor curved．However，they set，as outliersdistortthereal structureofthe datamanifold．Tosolvethe easily topological underlying problem， laws．It therelativetransformationonthe data is the relative originalspaceproposedbymodelingcognitive is thattherelativetransformationisakindofnonlinear transformation80thatitmakesthe proved enlarging transformeddatamore relativetransformationcanweakentheinfluenceof distinguishable．Meanwhile，the noiseondataandmakedatarelativedenser．Tomeasurethe anddistancebetweendata in similarity points with bethen to the relativeismoreconsistenttheintuitionof can space people，which appliedimprove machine