Design, Implementation and Analysis of a Parallel and Scalable Cascade Support Vector Machine Framework

Abstract

Natural sciences and engineering are experiencing a growing influx of large datasets that are analyzed using machine learning methods, both supervised and unsupervised. This in turn calls for scalable solutions that take advantage of parallel computing systems such as large supercomputing resources. Cascade support vector machine(CSVM) training offers a substantial boost to performance when processing large machine learning training data compared to more traditional parallel methods, with minimal loss of accuracy.This paper describes the design and implementation of a CSVM system that attempts to be highly parallel and provide a reasonable number of support vector machine(SVM) features. It also provides far better scaling than common parallel methods.The proposed approach attempts to keep CPU utilization high by mixing cascade and traditional parallel training methods, utilizing otherwise unused threads as workers in the more traditional parallel approach as they are freed up by the cascade method.Results indicate a 2.7 fold speedup in some tested datasets, improved scaling, and over 80\% reduced memory footprint compared to traditional parallel methods.

Raun- og nátturuvísindasviðin eru að kljást við sífellt fleiri stór gagnasett sem eru best leyst með vélnámi.Þetta kallar á vélnámsaðferðir sem eru skilvirkari, og nýta sér samhliðavinnslu í kerfum eins og ofurtölvum til þess að ná fram nægjanlegum vinnsluhraða.Þessi ritgerð lýsir hönnun á „Cascade SVM`` vélnámskerfi, sem er ætlað að bæta vinnslutíma á stórum gagnasettum á kostnað smávægilegrar nákvæmni.Hönnunin tekur líka til greina núverandi vélnámskerfi, og býður upp á svipað viðmót og svipaða valmöguleika.Aðferðin sem lýst er í þessari ritgerð hefur sýnt allt að 2.7 falda hraðaaukningu, mun betri skölun, og yfir 80\% minni minnisnotkun miðað við sambærilegar aðferðir.