On-line Learning

Main.On-lineLearning History

Hide minor edits - Show changes to output

April 21, 2009, at 11:15 AM by Zhdanov - lecture available
Changed lines 6-8 from:
* Michael Collins, Amir Globerson, Terry Koo, Xavier Carreras, and Peter Bartlett.
Exponentiated Gradient Algorithms for Conditional Random Fields and Max-Margin Markov Networks.
To appear in JMLR.
to:
* Michael Collins, Amir Globerson, Terry Koo, Xavier Carreras, and Peter Bartlett. Exponentiated Gradient Algorithms for Conditional Random Fields and Max-Margin Markov Networks. The Journal of Machine Learning Research, 9, pp. 1775-1822 2008.
April 21, 2009, at 11:13 AM by Zhdanov - lectures available
Changed line 3 from:
Several algorithms base on learning a perceptron $w \cdot x$ and get a bound on the number of mistakes in comparison with the optimization function of SVM. These algorithms use different updates for $w$. The most successful use the potential approach, when the update is represented as a gradient of a certain potential function. The [[Exponentiated Gradient]], for example, uses exponential potential function.
to:
Several algorithms base on learning a perceptron $w \cdot x$ and get a bound on the number of mistakes in comparison with the optimization function of SVM. These algorithms use different updates for $w$. The most successful use the potential approach, when the update is represented as a gradient of a certain potential function. The [[Exponentiated Gradient]], for example, uses exponential potential function. [[Winnow algorithm]] allows to get other bounds.
April 21, 2009, at 11:12 AM by Zhdanov - lecture available
Changed lines 1-4 from:
On-line learning is an approach to solve different machine learning tasks using algorithms of [[competitive on-line prediction]]. One of the most popular tasks is to train on-line a maximum-margin classifier, like SVM, based on a maximization of a convex function. The [[Exponentiated Gradient]] and [[Gradient Descent]] were somehow applied to find such a function in a dual or primal form.
Lectures
of Nicolo Cesa-Bianchi about online learning are accessible [[http://videolectures.net/mlss07_bianchi_onlle/ | here]]. Lectures of Manfred K. Warmuth about the role of [[http://en.wikipedia.org/wiki/Bregman_divergence | Bregman Divergences]] (a very powerful proof technique) in online learning are accessible [[http://videolectures.net/mlss06tw_warmuth_olbd/ | here]].

Several algorithms base on learning a perceptron $w \cdot x$ and get a bound on the number of mistakes in comparison with the optimization function of SVM. These algorithms use different updates for $w$. The most successful use the
potential approach, when the update is represented as a gradient of a certain potential function.
to:
On-line learning is an approach to solve different machine learning tasks using algorithms of [[competitive on-line prediction]]. One of the most popular tasks is to train on-line a maximum-margin classifier, like SVM, based on a maximization of a convex function. Lectures of Nicolo Cesa-Bianchi about online learning are accessible [[http://videolectures.net/mlss07_bianchi_onlle/ | here]]. Lectures of Manfred K. Warmuth about the role of [[http://en.wikipedia.org/wiki/Bregman_divergence | Bregman Divergences]] (a very powerful proof technique) in online learning are accessible [[http://videolectures.net/mlss06tw_warmuth_olbd/ | here]].

Several algorithms base on learning a perceptron $w \cdot x$ and get a bound on the number of mistakes in comparison with the optimization function of SVM. These algorithms use different updates for $w$. The most successful use the potential approach, when the update is represented as a gradient of a certain
potential function. The [[Exponentiated Gradient]], for example, uses exponential potential function.
April 21, 2009, at 10:57 AM by Zhdanov - lectures acessible
Added lines 3-4:

Several algorithms base on learning a perceptron $w \cdot x$ and get a bound on the number of mistakes in comparison with the optimization function of SVM. These algorithms use different updates for $w$. The most successful use the potential approach, when the update is represented as a gradient of a certain potential function.
April 21, 2009, at 10:42 AM by Zhdanov - lectures accessible
Changed line 2 from:
Lectures of Nicolo Cesa-Bianchi about online learning are accessible [[http://videolectures.net/mlss07_bianchi_onlle/ | here]]. Lectures of Manfred K. Warmuth about the role of Bregman Divergences (a very powerful proof technique) in online learning are accessible [[http://videolectures.net/mlss06tw_warmuth_olbd/ | here]].
to:
Lectures of Nicolo Cesa-Bianchi about online learning are accessible [[http://videolectures.net/mlss07_bianchi_onlle/ | here]]. Lectures of Manfred K. Warmuth about the role of [[http://en.wikipedia.org/wiki/Bregman_divergence | Bregman Divergences]] (a very powerful proof technique) in online learning are accessible [[http://videolectures.net/mlss06tw_warmuth_olbd/ | here]].
April 21, 2009, at 10:38 AM by Zhdanov - lectures accessible
Changed line 2 from:
Lectures of Nicolo Cesa-Bianchi about online learning are accessible [[http://videolectures.net/mlss07_bianchi_onlle/ | here]]. Lectures of Manfred K. Warmuth about the role of Bregman Divergences in online learning are accessible [[http://videolectures.net/mlss06tw_warmuth_olbd/ | here]].
to:
Lectures of Nicolo Cesa-Bianchi about online learning are accessible [[http://videolectures.net/mlss07_bianchi_onlle/ | here]]. Lectures of Manfred K. Warmuth about the role of Bregman Divergences (a very powerful proof technique) in online learning are accessible [[http://videolectures.net/mlss06tw_warmuth_olbd/ | here]].
April 21, 2009, at 10:35 AM by Zhdanov - indroductory lecture
Added line 2:
Lectures of Nicolo Cesa-Bianchi about online learning are accessible [[http://videolectures.net/mlss07_bianchi_onlle/ | here]]. Lectures of Manfred K. Warmuth about the role of Bregman Divergences in online learning are accessible [[http://videolectures.net/mlss06tw_warmuth_olbd/ | here]].
July 10, 2008, at 09:11 PM by Zhdanov - learn -- >find
Changed line 1 from:
On-line learning is an approach to solve different machine learning tasks using algorithms of [[competitive on-line prediction]]. One of the most popular tasks is to train on-line a maximum-margin classifier, like SVM, based on a maximization of a convex function. The [[Exponentiated Gradient]] and [[Gradient Descent]] were somehow applied to learn such a function in a dual or primal form.
to:
On-line learning is an approach to solve different machine learning tasks using algorithms of [[competitive on-line prediction]]. One of the most popular tasks is to train on-line a maximum-margin classifier, like SVM, based on a maximization of a convex function. The [[Exponentiated Gradient]] and [[Gradient Descent]] were somehow applied to find such a function in a dual or primal form.
Changed line 1 from:
On-line learning is an approach to solve different machine learning tasks using algorithms of [[competitive on-line prediction]]. One of the most popular tasks is to learn a maximum-margin classifier, like SVM, based on a maximization of a convex function. The [[Exponentiated Gradient]] and [[Gradient Descent]] were somehow applied to learn such a function in a dual or primal form.
to:
On-line learning is an approach to solve different machine learning tasks using algorithms of [[competitive on-line prediction]]. One of the most popular tasks is to train on-line a maximum-margin classifier, like SVM, based on a maximization of a convex function. The [[Exponentiated Gradient]] and [[Gradient Descent]] were somehow applied to learn such a function in a dual or primal form.
Added lines 1-7:
On-line learning is an approach to solve different machine learning tasks using algorithms of [[competitive on-line prediction]]. One of the most popular tasks is to learn a maximum-margin classifier, like SVM, based on a maximization of a convex function. The [[Exponentiated Gradient]] and [[Gradient Descent]] were somehow applied to learn such a function in a dual or primal form.

!!!Bibliography
* Michael Collins, Amir Globerson, Terry Koo, Xavier Carreras, and Peter Bartlett.
Exponentiated Gradient Algorithms for Conditional Random Fields and Max-Margin Markov Networks.
To appear in JMLR.
* Martin Zinkevich. Online convex programming and generalized infinitesimal gradient ascent. In ''Twentieth International Conference on Machine Learning'', 2003.