文档介绍：arXiv: [q-] 7 Nov 2008
ULS(VDA r usatal atr(pt 2tms th Librar times) Scientific (GNU 12 CBLAS to on (up based versions faster versio optimized substantially optimized CPU parallel are The GPU (NVIDIA) the algorithms. CUBLAS that GS-PCA par show Unit) and results NIPALS-PCA Processing merical (Graphics both GPU of the plementation discuss in we this Also, or eliminates Gram-Schmidt PCA. which fir on GS-PCA), the based (called algorithm of an ization estimation present the we Here to and nents. limited a orthogonality, algor usually mon NIPALS-PCA is loss standard the applicability from the sets suffers on unfortunately data based which is putation For PCA to analysis. data multivariate
rnia opnn nlss(C)i e ttsia tec statistical key a is (PCA) ponent Principal
aallGUIpeetto of Implementation GPU Parallel
nttt o icmlxt n Informatics and plexity for Institute
trtv C Algorithms PCA Iterative
50Uiest rv W Calgary NW, Drive University 2500
let,TN14 Canada 1N4, T2N Alberta,
nvriyo Calgary of University
.Andrecut M.
Abstract
1
po- few st
fNIPALS- of g
hrfr it’s therefore
s ae on based ns,
nqefor hnique
thogonal-
le im- allel
pproach
y).
nthe an
ithm,
nu-
1 Introduction
ponent analysis (PCA) is one of the most valuable results from ap-
plied linear algebra, and probably the most popular method used pacting
higher dimensional data sets into lower dimensional ones for data analysis, visu-
alization, feature extraction, or pression [Jackson, 1991, Jolliffe, 2002].
PCA provides a statistically optimal way of dimensionality reduction by pro-
jecting the data onto a lower-dimensional orthogonal subspace that captures
as much of the variation of the data as possible. Unfortunately, PCA quickly
es quite expensive pute for high-dimensional data sets, where both
the number of variables and samples is high. Therefore, there is a real need in
many applications to accelerate putation speed of PCA algorithms. For
large d