Difference between revisions of "MLReadingGroup"

Revision as of 13:46, 3 March 2008

Machine Learning Reading Group

Welcome to the wiki of the machine learning reading group.

Mailing list

We maintain an announcement/discussion list for the reading group. You may sign up for the list here.

Schedule (Spring 2008)

Our weekly meetings are Tue 1:00-2:30pm in the AI lab on the 4th floor of the CS building (CS 431).

Schedule of topics:

• 04 March 2008
  • Leader: Jordan Boyd-Graber
  • Paper: Toutanova, Kristina and Johnson, Mark. A Bayesian LDA-based model for semi-supervised part-of-speech tagging. (2007)
  • Paper: (If there's time, an additional paper)

• 26 February 2008
  • Leader: Berk Kapicioglu
  • Paper: Gilles Celeux, Didier Chauveau, Jean Diebolt. "On Stochastic Versions of the EM Algorithm.", 1995

Schedule (Fall 2007)

Our weekly meetings are Wed 4:00-5:30pm in the AI lab on the 4th floor of the CS building (CS 431).

Schedule of topics:

• 12 December 2007
  • Leader: Zafer Barutcuoglu
  • Paper: G.E. Hinton, S. Osindero, Y.W. Teh. "A Fast Learning Algorithm for Deep Belief Nets." Neural Computation, 2006.
  • More empirical results: Bengio et al. "Greedy Layer-wise Training of Deep Networks." NIPS, 2006.
• 28 November 2007
  • Leader: Umar Syed
  • Paper: Umar Syed and Robert E. Schapire. "A game-theoretic approach to apprenticeship learning", NIPS (2008).
  • Background reading: The work is based on Yoav Freund and Robert E. Schapire, "Game theory, on-line prediction, and boosting", COLT (1996) (see Section 2 and the Appendix).
• 14 November 2007
  • Leader: Melissa Carroll
  • Paper: Hui Zou and Trevor Hastie. "Regularization and variable selection via the elastic net." (2005) J. R. Statist. Soc. B, 67(2), pp. 301–320.
  • Background reading: It may be helpful to read up on or review LASSO and LARS. See the LASSO Page

• 24 October 2007
  • Patrón: Berk Kapicioglu
  • Paper: Joshua Goodman. "Exponential Priors for Maximum Entropy Models." North American ACL 2004.

Additional topics:

• relational network models
• DP + parse trees
• online learning
• semi-supervised learning
• stochastic gradient
• convex optimizing
• parallel learning
• game theory

Schedule (Spring 2007)

Our weekly meetings are Thu 1:30-3:00pm in the AI lab on the 4th floor of the CS building (CS 431).

Schedule of topics:

• 29 March 2007
  • Leader: Indraneel Mukherjee
  • Paper: Large Margin Hidden Markov Models for Automatic Speech Recognition
• 15 March 2007
  • Leader: Jordan Boyd-Graber
  • Paper 1: Catching the Drift: Probabilistic Content Models, with Applications to Generation and Summarization (2004)
  • Paper 2: Unsupervised topic modelling for multi-party spoken discourse (2006)
• 8 March 2007
  • Leader: Umar Syed
  • Paper: Pieter Abbeel and Andrew Y. Ng. "Apprenticeship learning via inverse reinforcement learning." ICML 2004.
• 1 March 2007
  • Leader: Miro Dudik
  • Paper: Shai Shalev-Shwartz and Yoram Singer. "Convex Repeated Games and Fenchel Duality."
  • See also a more recent version from NIPS 2006. It contains more references and the math is slightly different; e.g., it introduces strong convexity relative to a norm. I find it a little bit too condensed and more difficult to read.
• 22 February 2007
  • Leader: Melissa Carroll
  • Paper: R.A. Hutchinson, T. Mitchell, I. Rustandi. "Hidden Process Models." ICML 2006.
  • Background on fMRI classification application: T. Mitchell, R. Hutchinson, R. Niculescu, F. Pereira, X. Wang. "Learning to Decode Cognitive States from Brain Images." Machine Learning, 57, 145–175, 2004.
• 15 February 2007
  • Leader: Edo Airoldi
  • Paper: J. Lasserre, C. M. Bishop, and T. Minka. "Principled hybrids of generative and discriminative models." CVPR 2006.
  • Notes: T. Minka. "Discriminative models, not discriminative training." MSR-TR-144 2005.
• 8 February 2007
  • Leader: Joe Calandrino
  • Paper: I. Dinur, K. Nissim. "Revealing Information while Preserving Privacy." PODS 2003.
• inverse RL (Umar)
• disc/gen approaches (Bishop...)
• active learning (BK)
• hidden process models (MC)
• Gaussian processes (Z)
• Dirichlet processes
• Semisupervised learning (Florian)
  • The use of unlabelled data in predictive modelling, Liang F, Mukherjee S and West M, Statistical Science, to appear.
  • some paper by Lafferty and Wasserman ?
• Quantam neural networks (Vaneet)
• Manifold learning (Z)
• On-line learning (Berk)
• Music stuff/transcription (R)
• Variational methods (JC)
• Random projections (Charikar)

Schedule (Fall 2006)

Our weekly meetings are Fridays, 3pm to 5pm, in CS 402.

Scheduled readings:

• 5 October 2006 THURSDAY 4:30PM this week
  • Leader: Zafer
  • Paper: Horst, R., Thoai, N. V. 1999. DC Programming: Overview. J. Optim. Theory Appl.

  (If you need a review: Hindi, H. 2004. A Tutorial on Convex Optimization. American Control Conference.)

  (Application for the interested: Argyriou A. et al. 2006. A DC-Programming Algorithm for Kernel Selection. ICML.)

  And another application Ellis, S. and Nayakkankuppam, V. Phylogenetic Analysis Via DC Programming .

• 13 October 2006
  • Leader: Miro
  • Papers:
    • Robert Tibshirani. Regression shrinkage and selection via the Lasso. J. R. Statist. Soc. B, 1995.
    • Su-In Lee, Honglak Lee, Pieter Abbeel and Andrew Y. Ng. Efficient L1 regularized logistic regression. In Proceedings of the Twenty-First National Conference on Artificial Intelligence (AAAI-06), 2006.
  • Paper:
• 20 October 2006
  • Leader: Florian
  • Paper: N Meinshausen and P Bühlmann, High Dimensional Graphs and Variable Selection With the Lasso, Annals of Statistics 34(3), 1436-1462
  • Background reading: D Heckerman, DM Chickering, C Meek, R Rounthwaite, C Kadie, Dependency Networks for Inference, Collaborative Filtering, and Data Visualization, JMLR, 1(Oct):49-75, 2000
• 27 October 2006
  • Leader:
  • Paper:
• 3 November 2006
  • Leader:
  • Paper:
• 10 November 2006
  • "Leader": Jordan
  • Paper: Understanding the Yarowsky Algorithm Abney, S. 2004. Understanding the Yarowsky Algorithm. Comput. Linguist. 30, 3 (Sep. 2004), 365-395.
  • Background: The original paper Yarowsky, D. 1995. Unsupervised word sense disambiguation rivaling supervised methods. In Proceedings of the 33rd Annual Meeting on Association For Computational Linguistics (Cambridge, Massachusetts, June 26 - 30, 1995). Annual Meeting of the ACL. Association for Computational Linguistics, Morristown, NJ, 189-196.
  • Background: A good overview of the problem area Ide, N. and Véronis, J. 1998. Introduction to the special issue on word sense disambiguation: the state of the art. Comput. Linguist. 24, 1 (Mar. 1998), 2-40.
• 17 November 2006
  • Boss: Berk Kapicioglu
  • Paper: Weighted One-Against-All
  • Nostalgia: Rifkin and Klautau. "In Defense of One-Vs-All Classification." Journal of Machine Learning Research, Volume 5, pp. 101-141, 2004.
• 1 December 2006
  • Leader: Jonathan Chang
  • Paper: Snow, Jurafsky, and Ng. "Semantic Taxonomy Induction from Heterogenous Evidence." Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL, pages 801-808, July 2006.
  • Some useful (perhaps) background on the technique: Snow, Jurafsky, and Ng. "Learning Syntactic Patterns for Automatic Hypernym Discovery."
  • Background on what they're trying to do: Miller. "WordNet: A Lexical Database for English." Communications of the ACM, Volume 38, Issue 11 (November 1995), Pages 39-41.
• 15 December 2006
  • Leader: Rebecca Fiebrink
  • Paper: Kohavi, R. 1995. "A study of cross-validation and bootstrap for accuracy estimation and model selection." Proceedings of the International Joint Conferences on Artificial Intelligence (IJCAI).

Proposed Topics and Papers

Please add further topics, suggest papers for particular topics, etc. here.

• Lasso
  • The original paper: Robert Tibshirani. Regression shrinkage and selection via the Lasso. J. R. Statist. Soc. B 58, 1995.
  • The Lasso Page
  • Generalization properties:
    • Sara A. van de Geer. High-dimensional generalized linear models and the Lasso. Tech. Report, June 2006.
    • Keith Knight and Wenjiang Fu. Asymptotics for lasso-type estimators. Ann. Statist. 28, 2000.
  • Sparse approximation:
    • David Donoho. For most large underdetermined systems of linear equations, the minimal l1-norm near-solution approximates the sparsest near-solution, Tech. Report, August 2004.
    • David Donoho. For most large underdetermined systems of linear equations, the minimal l1-norm solution is also the sparsest solution, Tech. Report, September 2004.
  • Model selection: P. Buhlmann and B. Yu. Boosting, Model Selection, Lasso and Nonnegative Garotte. Tech. Report, 2005.
  • Relatives:
    • RODEO
    • Least angle regression
    • Hui Zou and Trevor Hastie. Regularization and Variable Selection via the Elastic Net. J. R. Statist. Soc. B, 2005 + Addendum
    • Su-In Lee, Honglak Lee, Pieter Abbeel and Andrew Y. Ng. Efficient L1 regularized logistic regression. In Proceedings of the Twenty-First National Conference on Artificial Intelligence (AAAI-06), 2006.
• Optimization
• Language Applications
• Music
  • Cemgil, A. T., H. J. Kappen, and D. Barber. 2005. A generative model for music transcription. IEEE Transactions on Speech and Audio Processing.
• Markov Decision Processes (MDPs) and Reinforcement Learning
  • State-space abstraction/aggregation in MDPs
  • The E^3 algorithm (Kearns and Singh)
  • Inverse reinforcement learning
• Active Learning
• Deep Neural Networks
• Factor Graphs
  Algorithms that must deal with complicated global functions of many variables often exploit the manner in which the given functions factor as a product of "local" functions, each of which depends on a subset of the variables.
  • Tutorial: Factor Graphs and the Sum-Product Algorithm, F.R. Kschischang, B. Frey, H-A Loelinger, IEEE Transactions on Information Theory, Vol 47, No 2, Feb 2001.
  • Application: Physical Network Models, C-H Yeang, T Ideker, T Jaakkola, Journal of Computational Biology, Vol 11, No 2-3, 2004
• Cross-validation
  • Kohavi, R. 1995. "A study of cross-validation and bootstrap for accuracy estimation and model selection." Proceedings of the International Joint Conferences on Artificial Intelligence (IJCAI).
    • "Do 10-fold CV instead of leave-one-out," at least when doing model selection
  • Rivals, I., and L. Personnaz. 1999. "On cross-validation for model selection." Neural Computation 11 (4).
    • "Do statistical tests instead of leave-one-out," at least when doing model selection
  • Elisseeff, A., and M. Pontil. 2003. "Leave-one-out error and stability of learning algorithms with applications." Advances in Learning Theory: Methods, Models and Applications, NATO Science Series III: Computer and Systems Sciences, Vol. 190, J. Suykens et al. Eds.
    • Reasons about sufficient conditions for LOO-CV error to approach generalization error for fixed algorithms (including kernel methods). A more theoretical (and maybe less practical) paper.
    • see also Evgeniou, T., M. Pontil, and A. Elisseeff. 2004. "Leave-one-out error, stability, and generalization of voting combinations of classifiers." Machine Learning 55:(1): 71-97. (I haven't read this but it's very related.)

Participants

(Participants, please add your name to the list below.)

Faculty

• David Blei
• Rob Schapire

PostDocs

• Edo Airoldi, LSI & CS
• Florian Markowetz, LSI

Students

• Indraneel Mukherjee, CS
• Zafer Barutcuoglu, CS
• Jordan Boyd-Graber, CS
• Joseph Calandrino, CS
• Melissa Carroll, CS
• Jonathan Chang, EE
• Miroslav Dudik, CS
• Rebecca Fiebrink, CS
• Berk Kapicioglu, CS
• Umar Syed, CS