
4th Annual Machine Learning Symposium
Friday, November 6, 2009
This is the fourth symposium on Machine Learning at the New York Academy of Sciences. The aim of these series of symposia is to build a community of scientists in machine learning from the NYC area's academic, government, and industrial institutions by convening and promoting the exchange of ideas in a neutral setting.
Speakers
- Robert Bell, AT&T Labs Research
- Avrim Blum, Carnegie Mellon University
- Thorsten Joachim, Cornell University
- Phil Long, Google
Presented by
Past Machine Learning Symposia
- 3rd Annual Machine Learning Symposium 2008
- 2nd Machine Learning Symposium 2007
- 1st Machine Learning Symposium 2006
Agenda
9:30 AM | Coffee & Poster Set-Up |
10:00 AM | Opening Remarks |
10:15 AM | Lessons from the Netflix Prize |
11:00 AM | A New Theoretical Framework for Clustering |
11:45 AM | Graduate Student Talks Ensemble Nyström Method An Event-Sensitive Bandit Algorithm A New Model for Multiclass Boosting Interactively Optimizing Information Retrieval Systems as a Dueling Bandits Problem Finding Latent Sources in Recorded Music with a Shift-Invariant HDP |
12:45 PM | Lunch & Poster Session |
2:30 PM | Support Vector Machines for Predicting Structured Outputs |
3:15 PM | On Noise-tolerant Learning using Linear Classifiers |
4:00 PM | Student Award Winner Announcement & Closing Remarks |
Speakers
Speakers
Robert Bell, PhD
AT&T Labs Research
Robert Bell has been a member of the Statistics Research Department at AT&T Labs-Research since 1998. He previously spent about twenty years at RAND doing public policy analysis. His current research interests include machine learning methods, analysis of data from complex samples, and record linkage methods. He was a member of the team that won the Netflix Prize competition in 2009. He has served on the Fellows Committee of the American Statistical Association, the board of the National Institute of Statistical Sciences, the Committee on National Statistics, and several National Research Council advisory committees studying statistical issues from conduct of the decennial census to airline safety.
Avrim Blum, PhD
Carnegie Mellon University
Avrim Blum is a Professor of Computer Science at Carnegie Mellon University. His main research interests are in Machine Learning Theory, Approximation Algorithms, and Algorithmic Game Theory, and he is also known for his work in AI Planning. He has served as Program Chair for the IEEE Symposium on Foundations of Computer Science (FOCS) and the Conference on Learning Theory (COLT), as well as on the organizing committee for the National Academy of Sciences U.S. Frontiers of Science Symposium. He was recipient of the Sloan Fellowship and NSF National Young Investigator Awards, the ICML/COLT 10-year best paper award, and is a Fellow of the ACM.
Thorsten Joachims, PhD
Cornell University
Thorsten Joachims is an Associate Professor in the Department of Computer Science at Cornell University. In 2001, he finished his dissertation with the title "The Maximum-Margin Approach to Learning Text Classifiers: Methods, Theory, and Algorithms", advised by Prof. Katharina Morik at the University of Dortmund. From there he also received his Diplom in Computer Science in 1997 with a thesis on WebWatcher, a browsing assistant for the Web. From 1994 to 1996 he was a visiting scientist at Carnegie Mellon University with Prof. Tom Mitchell. His research interests center on a synthesis of theory and system building in the field of machine learning, with a focus on Support Vector Machines and machine learning with text. He authored the SVM-Light algorithm and software for support vector learning.
Philip Long, PhD
Phil Long has worked in the research unit at Google for the past four years. He did his Ph.D. at UC Santa Cruz, and postdocs at Technische Universitaet Graz and Duke. Then he joined the faculty at the National University of Singapore, followed by a stint at the Genome Institute of Singapore. Next, he went to the Center for Computational Learning Systems at Columbia, and after that he joined Google.
Steering Committee
Corinna Cortes, PhD
Google, Inc.
Tony Jebara, PhD
Columbia University
Michael Kearns, PhD
University of Pennsylvania
John Langford, PhD
Yahoo Research
Mehryar Mohri, PhD
Courant Institute of Mathematical Sciences
Robert Schapire, PhD
Princeton University
David Waltz, PhD
Columbia University
Abstracts
Lessons from the Netflix Prize
Robert Bell, PhD, AT&T Labs Research
In October 2006, the DVD rental company Netflix released more than 100 million user ratings of movies for a competition to predict users’ ratings based on prior ratings. One allure to data analysts around the world was a $1,000,000 prize for a team achieving a ten percent reduction in root mean squared prediction error relative to Netflix’s current algorithm. The size of the data (over 17,000 movies and 480,000 users) and the nature of human-movie interactions produced many modeling challenges. After describing some of the techniques in use and advances spurred by the competition, I will offer lessons and raise some questions about building and regularizing massive prediction models, the role of statistics versus computer science in such endeavors, and prizes as a way to advance science. This is joint work with Chris Volinsky and Yehuda Koren, current and former colleagues at AT&T Labs-Research.
A New Theoretical Framework for Clustering
Avrim Blum, PhD, Carnegie Mellon University
Problems of clustering data come up in many different areas throughout science. Theoretical treatments of clustering have generally been of two types: either on algorithms for (approximately) optimizing various distance-based quantities such as k-median, k-means, and min-sum objectives, or on clustering under probabilistic "generative model" assumptions such as mixtures of Gaussian or related distributions. In this work we propose a new approach to analyzing the problem of clustering. Building on models used in computational learning theory, we consider the goal of approximately recovering an unknown target clustering from given pairwise similarity information, without making generative-model assumptions about the data. Instead, we ask: what relations between the similarity information and the desired clustering are sufficient to cluster well -- these relations taking the role of the "concept class" in learning theory. We find that if we are willing to relax our goals a bit (for example, allow the algorithm to produce a hierarchical clustering that we will call successful if it contains a pruning that is close to the correct answer) then this leads to a number of interesting graph-theoretic and game-theoretic properties that are sufficient for an algorithm to succeed. We show we can also produce accurate clusterings under implicit assumptions made when considering approximation algorithms for distance-based objectives such as k-median, k-means, and min-sum. In fact, in this case we can do as well as if we had been able to approximate such objectives to values that are NP-hard to achieve. This talk is based on work joint with Maria-Florina Balcan, Anupam Gupta, and Santosh Vempala.
Support Vector Machines for Predicting Structured Outputs
Thorsten Joachims, PhD, Cornell University
Over the last decades, much of the research on discriminative learning has focused on problems like classification and regression, where the prediction is a single univariate variable. But what if we need to predict complex objects like trees, orderings, or alignments? Such problems arise, for example, when a natural language parser needs to predict the correct parse tree for a given sentence, when one needs to optimize a multivariate performance measure like the F1-score, or when predicting the alignment between two proteins. This talk discusses a support vector approach and algorithm for predicting such complex objects. It generalizes conventional classification SVMs to a large range of structured outputs and multivariate loss functions. While the resulting training problems have exponential size, there is a simple algorithm that allows training in polynomial (or in many cases linear) time. The algorithm is implemented in the SVM-Struct software and empirical results will be given for several examples.
On Noise-Tolerant Learning using Linear Classifiers
Phil Long, PhD, Google
This talk is about learning using linear hypotheses in the presence of noise, including the following topics: * New algorithms that tolerate a lot of "malicious noise" given constraints on a probability distribution generating the examples. * The ability of linear classifiers to approximate the Bayes optimal error rate for some tree-structured two-layer sources with the class designation at the root, the observed variables at the leaves, and some hidden variables in between. * Limitations on the noise-tolerance of some boosting algorithms based on convex optimization. (This is joint work with Nader Bshouty, Adam Klivans and Rocco Servedio.)
Travel & Lodging
Our Location
The New York Academy of Sciences
7 World Trade Center
250 Greenwich Street, 40th floor
New York, NY 10007-2157
212.298.8600
Hotels Near 7 World Trade Center
Recommended partner hotel:
Club Quarters, World Trade Center
140 Washington Street
New York, NY 10006
Phone: (212) 577-1133
Located on the south side of the World Trade Center, opposite Memorial Plaza, Club Quarters, 140 Washington Street, is just a short walk to our location. The New York Academy of Sciences is a part of the Club Quarters network. Please feel free to make accommodations on-line to save significantly on hotel costs.
Password: NYAS
Other hotels located near 7 WTC:
212.945.0100 |
212.693.2001 |
212.385.4900 |
212.693.2001 |
212.269.6400 |
212.742.0003 |
212.232.7700 |
212.747.1500 |
212.344.0800 |