國立虎尾科技大學 |

Scalable Optimization Methods for Machine Learning : = Structures, Properties and Applications.

Record Type:	Language materials, manuscript : Monograph/item
Title/Author:	Scalable Optimization Methods for Machine Learning :/
Reminder of title:	Structures, Properties and Applications.
Author:	Tao, Shaozhe.
Description:	1 online resource (189 pages)
Notes:	Source: Dissertation Abstracts International, Volume: 79-12(E), Section: B.
Contained By:	Dissertation Abstracts International79-12B(E).
Subject:	Industrial engineering. -
Online resource:	click for full text (PQDT)
ISBN:	9780438168909

Scalable Optimization Methods for Machine Learning : = Structures, Properties and Applications.
Tao, Shaozhe.

Scalable Optimization Methods for Machine Learning :Structures, Properties and Applications. - 1 online resource (189 pages)

Source: Dissertation Abstracts International, Volume: 79-12(E), Section: B.

Thesis (Ph.D.)--University of Minnesota, 2018.

Includes bibliographical references

Many problems in machine learning can be formulated using optimization models with constraints that are well structured. Driven in part by such applications, the need to solve very large scale optimization models is pushing the performance limits on traditional state-of-art methods. In this thesis, we conduct a systematic study on scalable optimization methods. Our investigations mainly cover three aspects: the role of special structures in the models, convergence properties of algorithms, and applications in machine learning.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2018

Mode of access: World Wide Web

ISBN: 9780438168909Subjects--Topical Terms:

679492
Industrial engineering.
Index Terms--Genre/Form:

554714
Electronic books.

Scalable Optimization Methods for Machine Learning : = Structures, Properties and Applications.
LDR:04181ntm a2200361Ki 4500 001 917465
005 20181012133446.5
006 m o u
007 cr mn||||a|a||
008 190606s2018 xx obm 000 0 eng d
020 $a 9780438168909
035 $a (MiAaPQ)AAI10824345
035 $a (MiAaPQ)umn:19218
035 $a AAI10824345
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Tao, Shaozhe. $3 1191520
245 1 0 $a Scalable Optimization Methods for Machine Learning : $b Structures, Properties and Applications.
264 0 $c 2018
300 $a 1 online resource (189 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertation Abstracts International, Volume: 79-12(E), Section: B.
500 $a Advisers: Shuzhong Zhang; Daniel Boley.
502 $a Thesis (Ph.D.)--University of Minnesota, 2018.
504 $a Includes bibliographical references
520 $a Many problems in machine learning can be formulated using optimization models with constraints that are well structured. Driven in part by such applications, the need to solve very large scale optimization models is pushing the performance limits on traditional state-of-art methods. In this thesis, we conduct a systematic study on scalable optimization methods. Our investigations mainly cover three aspects: the role of special structures in the models, convergence properties of algorithms, and applications in machine learning.
520 $a First, we study popular scalable methods on sparse structured models, including alternating direction method of multipliers, coordinate descent method, proximal gradient method and accelerated proximal gradient method. In contrast to many global convergence results in the literature, we are particularly interested in the local convergence behavior. We establish the local bounds on the LASSO model, showing their eventual local linear convergence. We show that all of the methods can be treated as some eigenvalue problems, and therefore a spectral analysis becomes applicable. We also observe that when initiated far from the solution, the spectral analysis implies that one possibly get a sequence of iterates that appears to stagnate, but is actually taking small constant steps toward the solution. Moreover, we illustrate how the unaccelerated proximal gradient method can sometimes be faster when the iterates get close enough to the solution, as compared to the accelerated proximal gradient method. A comprehensive comparison of all methods is presented.
520 $a Next we move on to group sparse structured model. We develop an inverse covariance estimator that can regularize for overlapping group sparsity, and provide better estimates, especially when the dimension size is much larger than the number of samples. Furthermore, we extend the estimator into a general setting that covers any convex differentiable functions with conic constraints. The general estimator can exploit the domain structure to reduce the computation cost. The designed Frank-Wolfe method can leverage the decomposition within group structure, hence speeding up computation. Simulations and applications using real data justify both stability and scalability of this estimator, as the results show noticeable improvement.
520 $a Finally, we explore a certain low-rank structure in tensor. We construct the connection between the low-rank property in tensor and the group sparsity in its factor matrices. This provides a way to find a low-rank tensor decomposition via a regularized multiconvex optimization. In our approach, no prior knowledge of tensor rank is assumed. We propose to apply block coordinate descent method, since each block update can be implemented efficiently. Consequently, we show that our approach can be used to solve the tensor low-rank completion problem as well.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2018
538 $a Mode of access: World Wide Web
650 4 $a Industrial engineering. $3 679492
655 7 $a Electronic books. $2 local $3 554714
690 $a 0546
710 2 $a ProQuest Information and Learning Co. $3 1178819
710 2 $a University of Minnesota. $b Industrial and Systems Engineering. $3 1180522
773 0 $t Dissertation Abstracts International $g 79-12B(E).
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=10824345 $z click for full text (PQDT)