Computer Science Machine

Phone: 914-***-****

E-mail: **@*****.***

Web: http://hunch.net/~jl/

Research Interests

I want to solve machine learning. In a solved form, machine learning is

automatic, robust, scalable, and easily used for many problems. In

the last several years, we have made great progress in this direction.

Education M.S./Ph.D. - Computer Science,, June 2000/2002. Advisors: and My thesis was on tight sample complexity bounds, used for evaluation and learning algorithm design. B.S./B.S. - Physics/Computer Science,, June 1997.Employment History

Microsoft Research, New York, NY May 2012-present Senior Research Scientist

Yahoo! Research, New York, NY June 2006-April 2012 Senior Research Scientist

TTI-Chicago, Chicago, IL September 2003 - May 2006

Research Assistant Professor

IBM, TJ Watson, Yorktown, NY September 2002 - August 2003

Herman Goldstine Fellow

University of Pennsylvania, Philadelphia, PA June 2002 - August 2002

Postdoc with Michael Kearns

Activities

Blog: Machine Learning (Theory) at http://hunch.net

Tutorial: at .

Tutorial: at and .

Workshop: Organizer at .

Tutorial: at .

Tutorial: at

Workshop: Organizer at

Tutorial: and

School: Organizer Chicago 2005

Workshop: Organizer (Ab)Use of Bounds

Workshop: Organizer, 2003

Tutorial: and

Service

Program Chair:

New York ML Symposium co-organizer:

Area chair/Senior PC: ICML 2004

Program committee: ICML 2003, 2005 & 2008,, 2005, 2007 2004, 2005 and others

Reviewing: 2001, 2002, 2003, 2005 & 2007, 2002,,,, JCSS, TCS, JACM, and others

Mentoring I have worked with many students over time. In all cases, I try to learn something from them as well.

(Postdoc UPenn) (phd student UC Berkeley) (Professor Carnegie Mellon) (Professor Georgia Tech) (Professor UMinnesota) (Professor Hendrix U) (Professor U Maryland) (Research Professor U New Mexico) (Professor UMinnesota) (Postdoc, Microsoft Research) (phd student Cornell) (Nokia) (Microsoft Research) (Microsoft Research) (Professor Stanford) (Yahoo! Research)Joseph O'Sullivan (ex-Google :-) (Professor UTAustin) (Professor UToronto) (Professor EPFL) (Professor Pomona)Alex Strehl (Facebook) (Professor UCLA) (JPL) (Sandia) (Professor UFF) (Y! Research) = on work which became part of their thesis

Select Publications,, and (Editors), Cambridge University Press, 2011.,, and, ICML-2006

[Pat Goldberg Best Paper Award in Computer Science, Electrical Engineering, and Mathematics],, and CAPTCHA: Using Hard AI Problems for Security, Vin de Silva and . . 290, pages 2319-2323, 2000

Papers, forthcoming,,, :, CoRR abs/1110.4198 (2011),,,, Bidding Bandits: Bounded Budget Exploration in Auctions

All Publications, appendedResearch Programs

Much of my work can be organized into coherent programs.

Creating and using tight sample complexity bounds for evaluation and machine learning algorithm design. This was my thesis work, and a small program of work by others has grown out of it.Learning Reductions. We transform complex learning problems into simple ones and then use good algorithms for the simple ones to solve the complex problems. Here we designed the method of analysis as well as the algorithms, and again a small program of work by others is growing out of it.Scalable learning. The goal here is to scale up learning algorithms in all ways. We have addressed scaling in the number of parameters, the number of examples, and the number of labels.Active Learning. We proved that active learning was possible in the same noisy settings as supervised learning and eventually refined the algorithm to efficiently use any supervised algorithm as an oracle.Contextual Bandits. In many real-life situations, we only get

feedback about choices taken rather than choices not taken, as is

assumed in normal supervised learning. This requires a ground-up

rethink of what machine learning means. We have developed evaluators

(equivalent to test sets in supervised learning), optmized the use of

exploration information, and developed exponentially faster algorithms. is an open source machine learning system which encodes some of the above. It is currently the most scalable public (generalized) linear learning algorithm by a wide margin.PatentsUS 2011/0110231 VOLUNTARY ADMISSION CONTROL FOR TRAFFIC YIELD MANAGEMENTUS 2012/0016642 CONTEXTUAL-BANDIT APPROACH TO PERSONALIZED NEWS ARTICLE RECOMMENDATIONUS 2010/0057546 SYSTEM AND METHOD FOR ONLINE ADVERTISING USING USER SOCIAL INFORMATIONUS 2010/0010891 METHODS FOR ADVERTISEMENT SLATE SELECTIONUS 2009/0265227 Methods for Advertisement Display Policy ExplorationUS 2005/0091524 Confidential fraud detection system and methodUS 2010/0131496 PREDICTIVE INDEXING FOR FAST SEARCH References (Phd advisor) *****+@**.***.*** *******@***.*****.*** ********@**.*********.***

********@**.****.*** *******@******.** *****@***.***

Citizenship

Awards

Publications, All,,,, :, AIStats 2012.,, and (Editors), Cambridge University Press, 2011.,,, :, CoRR abs/1110.4198 (2011),,, :, in Scaling Up Machine Learning, Cambridge University Press, 2011.,,, : Cloud Control: Voluntary Admission Control for Intranet Traffic Management, Information Systems and e-Business Management,,, :, [Best Paper Award],,,, :, : 19-26 (2011)

[Notable Paper Award]

: Efficient Exploration in Reinforcement Learning. Encyclopedia of Machine Learning 2010: 309-311

Open Problem:, COLT 2010.,,, :, Algorithmica 58(4): 990-1021 (2010). Conference version in .

Kilian Q. Weinberger, Anirban Dasgupta,,, Josh Attenberg:,,,,, Alexander L. Strehl:,,, and :, . Journal version: IEEE Transactions on Computers 58(5): 662-676 (2009),, :, . Journal version: 75(1): 78-89 (2009), James Petterson,,,, S. V. N. Vishwanathan:, Journal of Machine Learning Research 10: 2615-2637 (2009), James Petterson,,,, Alexander L. Strehl, Vishy Vishwanathan:, Journal of Machine Learning Research - Proceedings Track 5 : 496-503 (2009),, :, . Journal version: Journal of Machine Learning Research 10: 777-801 (2009),, :, 75(3): 297-325 (2009),,,,,, :, .

[Winner of an Outstanding Paper Award], Alexander L. Strehl, :,,, Alexander L. Strehl:,,,,,, :, COLT 2007. Journal version: Machine Learning 72(1-2): 139-153 (2008), : The Epoch-Greedy Algorithm for Multi-armed Bandits with Side Information. NIPS 2007

Alexander L. Strehl,, Eric Wiewiora,, Michael L. Littman: PAC model-free reinforcement learning, ICML 2006,, and, ICML-2006

[Pat Goldberg Best Paper Award in Computer Science, Electrical Engineering, and Mathematics],, : Outlier detection by active learning, KDD 2006,, : Predicting Conditional Quantiles via Reduction to Classification, UAI 2006,, Continuous Experts and the Binning Algorithm, COLT 2006, : A comparison of tight generalization error bounds. ICML 2005

and . Journal version: Machine Learning 66(2-3): 119-149 (2007),, and CAPTCHA: Using Hard AI Problems for Security, and : Telling humans and computers apart automatically. Commun. ACM 47(2): 56-60 (2004)

John Langford: Combining Trainig Set and Test Set Bounds. ICML 2002: 331-338,, Competitive Analysis of the Explore/Exploit Tradeoff

(at xxx.lanl.gov and ) May 2002

and, (Not) Bounding the True Error, Vin de Silva and . . 290, pages 2319-2323, 2000

and . . and 5: 529-547 (2004)

Joseph O'Sullivan,, and . FeatureBoost: A Meta-Learning Algorithm that Improves Model Robustness.

and 1999. . also, 51(2): 165-179 (2003),, and 1999. Beating the Holdout: Bounds for KFold and Progressive Cross-Validation.,, and 1999. Monte Carlo Hidden Markov Models: Learning Non-Parametric Models of Partially Observable Stochastic Proecesses.

and :, .,, and :, .

Contact this candidate