on-line learning | Machine Learning Reading Group

Sent September 9:

This week we will go over the paper:

Lin Xiao, Dual Averaging Methods for Regularized Stochastic Learning and Online Optimization, in JMLR 2010

(NIPS version http://machinelearning.wustl.edu/mlpapers/paper_files/NIPS2009_1048.pdf)

It talks about primal-dual averaging, one of the methods mentioned in the review we covered which we didn’t quite work out.

Sent August 26:

Reminder – As we discussed last week, we will talk about the second part of the optimization lecture on Thursday.

The slides: http://pages.cs.wisc.edu/~swright/nips2010/sjw-nips10.pdf

The video: http://videolectures.net/nips2010_wright_oaml/

Sent August 13:

The paper for this week is “Stochastic Dual Coordinate Ascent Methods for Regularized Loss Minimization”.

Sent August 5:

I want to dedicate a few sessions to the area of optimization in ML. The idea is to cover new results but also try to make a “map” of the area, and make the connections between the fields.

To bring us all up to some level, this week, instead of reading a paper on a specific algorithmic/theoretical result, I thought we should read a review on the subject. I couldn’t find a good one written but found a nice nips tutorial on the subject (so you don’t even have to read:)).

The slides: http://pages.cs.wisc.edu/~swright/nips2010/sjw-nips10.pdf

The video: http://videolectures.net/nips2010_wright_oaml/

Sent July 28:

This week we will read the paper “Sparse coding for multitask and transfer learning”

from ICML 2013.

Sent July 14:

This week (Thursday @14:30) we will continue with Gaussian Processes. The subject will be the paper from ICML2011: http://www.icml-2011.org/papers/323_icmlpaper.pdf, which applies GPs to Reinforcement Learning.

Also, here a motivational video (learning this task previously demanded hundreds of trials, this algorithm does it in 7):

http://www.youtube.com/watch?v=XiigTGKZfks

Sent July 3:

It was suggested that we do a couple of sessions on Gaussian Processes. For next week, please read Chapters 2 and 5 of the book Gaussian Processes for Machine Learning, available at http://www.gaussianprocess.org/gpml/.

Sent June 16:

This week we will go over the paper:

“A Provably Efficient Algorithm for Training Deep Networks” http://arxiv.org/abs/1304.7045

Sent June 9:

Odalric will lead the discussion on the paper –

“Follow the Leader If You Can, Hedge If You Must”
http://arxiv.org/pdf/1301.0534v2.pdf, by Steven de Rooij, Tim van Erven, Peter D. Grünwald, Wouter M. Koolen.
This paper considers the online learning setting and tries to find a way to optimally
tune the Hedge algorithm so as to get an (really) adaptive algorithm.

A quick reference to the Hedge Algorithm: http://onlineprediction.net/n=Main.HedgeAlgorithm

Some motivation why this setting is useful, can be found in http://hal.archives-ouvertes.fr/docs/00/71/51/77/PDF/Devaine-Goude-Stoltz-Gaillard.pdf).

Sent April 29:

This week we’ll be reading the paper “Sparse inverse covariance estimation with the graphical lasso” by Friedman et al. (http://www-stat.stanford.edu/~tibs/ftp/graph.pdf). The paper discusses the problem of estimating sparse graphs by a lasso penalty applied to the inverse covariance matrix. The connection to graphs- conditional independence may be deduced when there is a zero in the inverse covariance matrix; for a reminder on the subject see the attached tutorial.

This week we will cover the paper –
Online Passive-Aggressive Algorithms
by Crammer et al. This group will be especially interesting thanks to Koby who has kindly agreed to share some insights into this work.

Abstract –
We present a family of margin based online learning algorithms for various prediction tasks. In particular we derive and analyze algorithms for binary and multiclass categorization, regression, uniclass prediction and sequence prediction. The update steps of our different algorithms are all based on analytical solutions to simple constrained optimization problems. This uniﬁed view allows us to prove worst-case loss bounds for the different algorithms and for the various decision problems based on a single lemma. Our bounds on the cumulative loss of the algorithms are relative to the smallest loss that can be attained by any ﬁxed hypothesis, and as such are applicable to both realizable and unrealizable settings. We demonstrate some of the merits of the proposed algorithms in a series of experiments with synthetic and real data sets.

We have moved the time back to 9:30.

See you on Wednesday (2/1/2012)!