A generalized online mirror descent with applications to classification and regression

被引:42
作者
Orabona, Francesco [1 ]
Crammer, Koby [2 ]
Cesa-Bianchi, Nicolo [3 ]
机构
[1] Yahoo Labs, New York, NY 10036 USA
[2] Technion Israel Inst Technol, Dept Elect Engn, IL-32000 Haifa, Israel
[3] Univ Milan, Dept Comp Sci, I-20135 Milan, Italy
关键词
RELATIVE LOSS BOUNDS; SUBGRADIENT METHODS;
D O I
10.1007/s10994-014-5474-8
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
Online learning algorithms are fast, memory-efficient, easy to implement, and applicable to many prediction problems, including classification, regression, and ranking. Several online algorithms were proposed in the past few decades, some based on additive updates, like the Perceptron, and some on multiplicative updates, like Winnow. A unifying perspective on the design and the analysis of online algorithms is provided by online mirror descent, a general prediction strategy from which most first-order algorithms can be obtained as special cases. We generalize online mirror descent to time-varying regularizers with generic updates. Unlike standard mirror descent, our more general formulation also captures second order algorithms, algorithms for composite losses and algorithms for adaptive filtering. Moreover, we recover, and sometimes improve, known regret bounds as special cases of our analysis using specific regularizers. Finally, we show the power of our approach by deriving a new second order algorithm with a regret bound invariant with respect to arbitrary rescalings of individual features.
引用
收藏
页码:411 / 435
页数:25
相关论文
共 48 条
[1]   Interior-Point Methods for Full-Information and Bandit Online Learning [J].
Abernethy, Jacob D. ;
Hazan, Elad ;
Rakhlin, Alexander .
IEEE TRANSACTIONS ON INFORMATION THEORY, 2012, 58 (07) :4164-4175
[2]  
[Anonymous], 2008, C LEARN THEOR COLT
[3]  
[Anonymous], 2007, NIPS
[4]  
[Anonymous], 2009, Advances in Neural Information Processing Systems
[5]  
[Anonymous], 1998, P 11 INT C NEUR INF
[6]  
[Anonymous], 1997, EL P 5 INT S ART INT
[7]   Adaptive and self-confident on-line learning algorithms [J].
Auer, P ;
Cesa-Bianchi, N ;
Gentile, C .
JOURNAL OF COMPUTER AND SYSTEM SCIENCES, 2002, 64 (01) :48-75
[8]   Relative loss bounds for on-line density estimation with the exponential family of distributions [J].
Azoury, KS ;
Warmuth, MK .
MACHINE LEARNING, 2001, 43 (03) :211-246
[9]  
Bauschke HH, 2011, CMS BOOKS MATH, P1, DOI 10.1007/978-1-4419-9467-7
[10]   Mirror descent and nonlinear projected subgradient methods for convex optimization [J].
Beck, A ;
Teboulle, M .
OPERATIONS RESEARCH LETTERS, 2003, 31 (03) :167-175