Big Data Optimization in Machine Learning / Xiaocheng Tang

Tang, Xiaocheng
Bib ID
vtls002096444
出版項
Ann Arbor, Michigan : ProQuest Information and learning, 2016.
稽核項
1 online resource (163 pages).
電子版
附註項
數位化論文典藏聯盟
預約人數:0
全部評等: 0
沒有紀錄。
 
 
 
03787ntm a2200505 i 4500
001
 
 
vtls002096444
003
 
 
VRT
005
 
 
20170512230800.0
006
 
 
m     o  d       
007
 
 
cr m|unnnup|||
008
 
 
170512s2016    miu     obm         eng d
020
$a 9781339483276 $q (ebook.)
035
$a (MiAaPQ)AAI10014082
039
9
$y 201705122308 $z VLOAD
040
$a MiAaPQ $b eng $e rda $c MiAaPQ $d TKU
100
1
$a Tang, Xiaocheng, $e author
245
1
0
$a Big Data Optimization in Machine Learning / $c Xiaocheng Tang
264
1
$a Ann Arbor, Michigan : $b ProQuest Information and learning, $c 2016.
264
4
$c ©2016
300
$a 1 online resource (163 pages).
336
$a text $b txt $2 rdacontent
337
$a computer $b c $2 rdamedia
338
$a online resource $b cr $2 rdacarrier
347
$a text file $b PDF $2 rda
490
1
$a Dissertation Abstracts International ; $v 77-07B(E)
500
$a Source: Dissertation Abstracts International, Volume: 77-07(E), Section: B.
500
$a Adviser: Katya Scheinberg.
502
$a Thesis $b (Ph.D.)-- $c Lehigh University, $d 2016
504
$a Includes bibliographical references
506
$a Access restricted to Tamkang University users.
520
$a Modern machine learning practices at the interface of big data, distributed environment and complex learning objectives post great challenges to designing scalable optimization algorithms with theoretical guarantees.  This thesis, built on the recent advances in randomized algorithms, concerns development of such methods in practice and the analysis of their theoretical implications in the context of large-scale structured learning problems, such as regularized regression/classification, matrix completion, hierarchical multi-label learning, etc. The first contribution of this work is thus a hybrid hierarchical learning system that achieve efficiency in a data-intensive environment. The intelligent decoding scheme inside the system further enhances the learning capacity by enabling a rich taxonomy representation to be induced in the label space. Important factors affecting the system scalability are studied and further generalized. This leads to the next contribution of the work -- a globally convergent inexact proximal quasi-Newton framework and the novel global convergence rate analysis. This work constitutes the first global convergence rate result for an algorithm that uses randomized coordinate descent to inexactly optimize subproblems at each iteration. The analysis quantifies precisely the complexity structure of proximal Newton-type algorithms, which makes it possible to optimize based on that structure to reduce complexity. The final contribution of the work is a practical algorithm which enjoys global convergence guarantee from the framework. The algorithm is memory- and communication-efficient and directly addresses the big data learning cases when both N (samples) and n (features) are large. We demonstrated that this general algorithm is very effective in practice and is competitive with state-of-the-art specialized methods.
533
$a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2016
538
$a Mode of access: World Wide Web
546
$a English
591
$a 數位化論文典藏聯盟 $b PQDT $c 淡江大學(2017)
653
$a Operations Research.
653
$a Computer Science.
653
$a Artificial Intelligence.
655
7
$a Electronic books. $2 local
700
1
$a Scheinberg, Katya, $e thesis advisor
710
2
$a ProQuest Information and Learning Co.
710
2
$a Lehigh University. $b Industrial Engineering.
830
0
$a Dissertation Abstracts International ; $v 77-07B(E).
856
4
1
$u http://info.lib.tku.edu.tw/ebook/redirect.asp?bibid=2096444 $z click for full text (PQDT)
999
$a VIRTUA00
沒有評論
叢書名
Dissertation Abstracts International ; 77-07B(E)
Dissertation Abstracts International ; 77-07B(E).
標題
摘要
Modern machine learning practices at the interface of big data, distributed environment and complex learning objectives post great challenges to designing scalable optimization algorithms with theoretical guarantees. This thesis, built on the recent advances in randomized algorithms, concerns development of such methods in practice and the analysis of their theoretical implications in the context of large-scale structured learning problems, such as regularized regression/classification, matrix completion, hierarchical multi-label learning, etc. The first contribution of this work is thus a hybrid hierarchical learning system that achieve efficiency in a data-intensive environment. The intelligent decoding scheme inside the system further enhances the learning capacity by enabling a rich taxonomy representation to be induced in the label space. Important factors affecting the system scalability are studied and further generalized. This leads to the next contribution of the work -- a globally convergent inexact proximal quasi-Newton framework and the novel global convergence rate analysis. This work constitutes the first global convergence rate result for an algorithm that uses randomized coordinate descent to inexactly optimize subproblems at each iteration. The analysis quantifies precisely the complexity structure of proximal Newton-type algorithms, which makes it possible to optimize based on that structure to reduce complexity. The final contribution of the work is a practical algorithm which enjoys global convergence guarantee from the framework. The algorithm is memory- and communication-efficient and directly addresses the big data learning cases when both N (samples) and n (features) are large. We demonstrated that this general algorithm is very effective in practice and is competitive with state-of-the-art specialized methods.
附註
Source: Dissertation Abstracts International, Volume: 77-07(E), Section: B.
Adviser: Katya Scheinberg.
Thesis
Includes bibliographical references
English
數位化論文典藏聯盟
合著者
ISBN/ISSN
9781339483276