18. Passive Aggressive (続)
wi+1
:=
wi
+
y
l(x,
y,
w)/(|x|2
+
1/C)
x
更更新式
l PAの最適化問題は閉じた解を持ち、次のように更更新可能
wi+1 := wi + αAx
l E
=
1
l α=
L(x,
y,
w)
/
(|x|2
+
1/C)
l A
=
I
l α∝L(x,
y,
w)であり、誤った割合に⽐比例例した更更新幅を使う
50. 出典
l [Collins EMNLP 02] “Discriminative Training Methods for Hidden
Markov Models: Theory and Experiments with Perceptron Algorithms.
Michael Collins, EMNLP 2002,
l [Crammer, JMLR 06] “Online Passive-Aggressive Algorithms”, Koby
Crammer, Ofer Dekel, Joseph Keshet, Shai Shalev-Shwartz, Yoram
Singer, Journal of Machine Learning, 2006
l [Dredze+, ICML 08] “Confidence-Weighted Linear Classification”,
Mark Dredze, Koby Crammer and Fernando Pereira , ICML 2008
l [Crammer+ NIPS 08] “Exact Convex Confidence-Weighted Learning”,
Koby Crammer, Mark Dredze and Fernando Pereira, NIPS 2008
51. l [Duchi+ 09] “Online and Batch Learning using Forward Backward
Splitting”, John Duchi and Yoram Singer, JMLR
l [S. S.-Shwartz 07] “Online Learning: Theory, Algorithms, and
Applications”, S. Shalev-Shwartz, Ph. D thesis 2007
l [Hazan+ ML 07] “Logarithmic regret algorithms for online convex
optimization”, E. Hazan, A. Agarwal and S. Kale. Machine Learning
2007
l [Hazan+ 11] “The convex optimization approach to regret
minimization”,
http://ie.technion.ac.il/~ehazan/papers/opt_book.pdf
52. l [Clarkson, FOCS 10] “Sublinear optimization for machine learning”,
K. L. Clarkson, E. Hazan and D. P Woodruff, FOCS 2010
.
l [Shalev+ ICML 07] “Pegasos: Primal estimated sub-gradient solver
for SVM”,S. Shalev-Shwartz and N. Srebro, ICML 2007
l [Hazan+ NIPS 11] “Beating SGD: Learning SVMs in Sublinear Time”,
E. Hazan, T. Koren, and N. Srebro, NIPS 2011