Vektormasinate paralleeliseerrimine
Date
2013
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Tartu Ülikool
Abstract
Tugivektormasin (Support Vector Machine) on masinõppe meetod, mida kasutakse andmete klassifitseerimiseks. Binaarse klassifikatsiooni probleem seisneb sellise funktsiooni või mudeli leidmisel, mis oskaks ennustada, mis klassi etteantud punkt x kuulub. Mudeli treenimiseks kasutatakse treeningandmeid.
Oma töös võrdlesime iteratiivsed ja paralleelseid tugivektormasina algoritmide implementatsioone.
Uurimise käigus avastasime et paralleelsed algoritmid, nagu oligi oodatud, töötavad palju kiiremini kui iteratiivsed, seejuures valesti klassifitseeritud
punktide arv ei suurene.
One of the techniques used for data classification is support vector machine (SVM). SVM takes binary classification as the fundamental problem and follows the geometrically intuitive approach to find a hyperplane that divides objects into two separate classes. The training part of the SVM aims to both maximize the width of the margin that surrounds the separating hyperplane and minimize the occurrence of classification errors. The goal of given thesis is to research efficiency in performance gained by using parallel approach to solve SVM and compare proposed techniques for parallelization in accuracy and computation speed.
One of the techniques used for data classification is support vector machine (SVM). SVM takes binary classification as the fundamental problem and follows the geometrically intuitive approach to find a hyperplane that divides objects into two separate classes. The training part of the SVM aims to both maximize the width of the margin that surrounds the separating hyperplane and minimize the occurrence of classification errors. The goal of given thesis is to research efficiency in performance gained by using parallel approach to solve SVM and compare proposed techniques for parallelization in accuracy and computation speed.