Boosted feature generation for classification problems involving high numbers of inputs and classes [Elektronische Ressource] / vorgelegt von Thomas Hörnlein

INAUGURAL—DISSERTATIONzurErlangungderDoktorwürdederNaturwissenschaftlich-MathematischenGesamtfakultätderRuprecht–Karls–UniversitätHeidelbergvorgelegtvonDipl.-Ing. ThomasHörnleinausHildburghausenTagdermündlichenPrüfung: 21.07.2010ThemaBoostedFeatureGenerationforClassificationProblemsInvolvingHighNumbersofInputsandClassesGutachter: Prof. Dr. BerndJähneProf. Dr. Dr. h. c. HansGeorgBockAbstractClassification problems involving high numbers of inputs and classes play animportant role in the field of machine learning. Image classification, in partic-ular, is a very active field of research with numerous applications. In additionto their high number, inputs of image classification problems often show sig-nificant correlation. Also, in proportion to the number of inputs, the numberof available training samples is usually low. Therefore techniques combininglow susceptibility to overfitting with good classification performance have tobe found. Since for many tasks data has to be processed in real time, computa-tionalefficiencyiscrucial aswell.Boostingisamachinelearningtechnique,whichisusedsuccessfullyinanum-ber of application areas, in particular in the field of machine vision. Due to it’smodulardesignandflexibility,Boostingcanbeadaptedtonewproblemseasily.In addition, techniques for optimizing classifiers produced by Boosting withrespect to computational efficiency exist. Boosting builds linear ensembles ofbaseclassifiersinastage-wisefashion.