February 2009 - talks - lists.cs.princeton.edu

Talk by John Langford next Monday, March 2 -- 1st speaker of the Machine Learning Speaker Series
by Melissa Lawson 25 Feb '09

25 Feb '09

This semester, the machine learning group will be organizing a speaker series, with the goal of inviting prominent researchers in machine learning to give talks at our department. The series is being sponsored by Yahoo!. Our first speaker is John Langford from Yahoo! Research, who will be visiting the department next Monday (March 2). John is an expert in the theory of machine learning. He also writes a popular blog about machine learning, http://hunch.net. John will be giving a talk at 4:25pm, in CS 105 (small auditorium). The title and abstract for his talk appear below. Future talk announcements for the speaker series will be sent to ML- stats-talk, which is a mailing list for announcing talks related to machine learning and statistics at Princeton. The list can be joined by following this link: https://lists.cs.princeton.edu/mailman/listinfo/ml-stat-talks If you'd like to meet John on Monday, details about his schedule will be sent to ML-stats-talk soon. Title: Use of Hash in Machine Learning Abstract: I'll describe a learning algorithm (Vowpal Wabbit) which effectively learns on and uses terascale datasets with a collection of tricks. One of the central tricks is rather fascinating---it uses a core hash representation which is both extremely time and space efficient. Stated another way, we use a sparsity preserving random projection, for which we prove convergence bounds reminiscent of the JL Lemma or locality sensitive hashing. This trick enables otherwise impossible applications like building a personalized spam filter for hundreds of thousands of users that easily fits into the RAM of a single machine.

1 0

A Golovinskiy preFPO
by Melissa Lawson 24 Feb '09

24 Feb '09

Alex Golovinskiy will present his preFPO on Wednesday March 4 at 11:30 AM in Room 401 (note room). The members of his committee are: Tom Funkhouser, advisor; Adam Finkelstein and Szymon Rusinkiewicz, readers; David Dobkin and Rob Schapire, nonreaders. Everyone is invited to attend his talk. His abstract follows below. ------------------------ Algorithms for analyzing the texture, symmetry, and part structure of 3D surface models Abstract: While much of computer graphics research has focused on low-level primitives, many computer graphics applications require reasoning about higher level structures. In response to this need, research has emerged to find and take advantage of higher level structures such as textural properties of geometry, symmetries, and functional parts. We advance this line of research in several directions. First, we present a method for analyzing and modifying detailed facial geometry. We introduce a statistical technique for the analysis and synthesis of small three-dimensional facial features, such as wrinkles and pores. As an initial step, our method separates the skin surface details from a smooth base mesh. Then, we analyze the resulting displacement maps and synthesize new geometry using tools from texture synthesis research. We demonstrate this method for analysis of changes of facial texture with respect to age and gender, detail-preserving interpolation between high-resolution face models, adding detail to low-resolution face models, and adjusting the apparent age of face models. Second, we present a framework for symmetry-aware mesh processing. Although perfect, partial, and approximate symmetries are pervasive in real-world geometry, current geometry processing algorithms ignore them. We present a framework in which, given a set of symmetries, we (i) warp the geometry to be symmetric, and (ii) re-mesh a model to have symmetric triangulation. We show how to use this framework to create algorithms that respect the symmetries of a model, and demonstrate applications for modeling, beautification, and simplification of nearly symmetric surfaces. Third, we present several investigations into segmenting meshes into parts. We present a hierarchical aggregation algorithm for mesh segmentation. We then expand this algorithm to consistently segment a set of meshes. We also show how to modify mesh segmentation algorithms to produce a soft, randomized version of segmentation. Finally, we present a framework for finding and recognizing objects in outdoor point cloud scans. We present methods for locating potential objects, segmenting them, and classifying them. We quantitatively evaluate these models on a part of an Ottawa scan.

1 0

FW: CS Events Calendar
by Ginny Hogan 18 Feb '09

18 Feb '09

_____ From: Laura Cummings-Abdo [mailto:lcumming@Princeton.EDU] Sent: Wednesday, February 18, 2009 10:34 AM To: Ginny Hogan Subject: CS Events Calendar Hi Ginny, Could you put this on the CS Events Calendar? Thanks. Symposium: Community and Municipal Wireless Systems Title: Community and Municipal Wireless Systems When: Thursday, March 5, 2009 - 4:30 PM - 6:30 PM Where: Woodrow Wilson School, Robertson Hall, Bowl 1 The Center is committed to bringing together technologists and policy scholars around critical issues in the regulation of technology and the use of technology to achieve policy ends. We have identified the complex set of issues around community and municipal wireless as an especially important topic, due to its potential to address issues of equity in access to the Internet and related technologies, as well as its potential to increase the effectiveness of municipal services. At the same time, we recognize that wireless projects have met with mixed success, and that both technical and political/economic issues remain to be resolved. We hope that this symposium will clarify the state of play, point to challenges and unanswered questions, and perhaps inspire some students and younger scholars to take a long-term interest in this area. Panelists: H. Vincent Poor, Dean of Princeton University's Engineering School Sharon Gillett, Commissioner of the State of Massachusetts Department of Telecommunications and Cable Greg Goldman, Chief Executive Officer, Wireless Philadelphia Sascha Meinrath, Research Director, New America Foundation Wireless Futures Program and Coordinator, Open Source Wireless Coalition http://citp.princeton.edu/events/community-wireless-panel/ Laura Cummings-Abdo Program Assistant Center for Information Technology Policy Princeton University Sherrerd Hall, Room 303 Princeton, NJ 08544 Tel: 609-258-9658 Fax: 609-964-1855 <http://citp.princeton.edu/> http://citp.princeton.edu/ Sherrerd-HallSmall

1 0

E Banks FPO
by Melissa Lawson 17 Feb '09

17 Feb '09

Eric Banks will present his FPO on Monday February 23 at 1:30PM in Room 402. The members of his committee are: Mona Singh, advisor; Olga Troyanskaya and Bernard Chazelle, readers; Tom Funkhouser and David Blei, nonreaders. A copy of his thesis, "Algorithms for Analyzing and Interrogating Protein Interaction Networks", is on display in room 310. Everyone is invited to attend his talk.

1 0

U Syed preFPO
by Melissa Lawson 13 Feb '09

13 Feb '09

Umar Syed will present his preFPO on Friday February 20 at 10:30AM in Room 302. The members of his committee are: Rob Schapire advisor; David Blei and Warren Powell (ORF), readers; Michael Littman (Rutgers) and Yael Niv (PSY), non-readers. Everyone is invited to attend his talk. His abstract follows below. --------------------------- Title: Algorithms for Hybrid Supervised/Reinforcement Learning Problems Abstract: Reinforcement and supervised learning are two different approaches to machine learning. For an illustration their differences, consider the task of learning to drive a car. A reinforcement learning algorithm learns how to drive be trying several different driving behaviors, and uses feedback from the environment to decided which behavior is best. The feedback is delivered in the form of rewards, which reflect the desirability of various events such as crashing, staying on the road, etc. By contrast, a supervised learning algorithm passively observes an expert driver, and attempts to imitate his driving behavior. In general, reinforcement learning algorithms are concerned with maximizing reward, and supervised learning algorithms are concerned with matching observed behavior. In this talk, I will discuss learning algorithms that combine aspects of both reinforcement and supervised learning, leveraging the strengths of each approach. Specifically, I will describe algorithms for (i) using observations from an expert to learn high-reward behavior in environments where the true rewards are only partially known; (ii) improving the imitation of an expert's behavior by imposing reward-based constraints on the space of possible behaviors; (iii) learning high-reward behavior in an environment where the rewards can abruptly, but predictably, change. I will present theoretical performance guarantees for each of our algorithms, which in some cases represent substantial improvements over guarantees for earlier methods. Our algorithms are motivated by problems in vehicle navigation, spoken dialog systems, and web search. I will present experimental results from some of these domains. Includes joint work with Rob Schapire, Michael Bowling (U Alberta), Jason Williams (AT&T), Nina Mishra (Microsoft), and Alex Slivkins (Microsoft).

1 0

J Boyd-Graber preFPO
by Melissa Lawson 13 Feb '09

13 Feb '09

Jordan Boyd-Graber will present his preFPO on Tuesday February 17 at 10:30 AM in Room 302. The members of his committee are: David Blei, advisor; Christiane Fellbaum and Ryan MacDonald (Google), readers; Rob Schapire and Dan Osherson (Psych), nonreaders. Everyone is invited to attend his talk. His abstract follows below. ----------------------------- Title: Linguistic Extensions to Topic Models Topic Models have been an active area of research in recent years, and models like latent semantic analysis, probabilistic latent semantic indexing, and latent Dirichlet allocation (LDA) have been successfully used for detecting opinions, finding similar images, and finding relevant documents given a query. However, such models make few assumptions about the input data: words are unrelated to each other and the words in a document are completely exchangeable (the so-called bag of words model). In this work, we present algorithms that enhance the document-level knowledge provided by LDA with richer linguistic assumptions. First, we allow topic models to use words arranged in a tree (such as the WordNet ontology) rather than a simple flat list and derive inference using MCMC for this model. One application that is made possible by this change is word sense disambiguation, which discovers the meanings of words (i.e. discriminating between "bank" the financial institution and the landform). We show that incorporating topics in this model improves disambiguation accuracy. Secondly, we present a model that incorporates local syntactic information into topic models, which allows the algorithm to find groups of words that are both globally thematically consistent and locally syntactically consistent. We use the product of experts model to combine document and syntactic information and derive variational inference procedures for this model. We show that this model predicts word usage better than previous models. Yours, Jordan

1 0

last reminder: CS Colloquium (TODAY): Stuart Geman]
by Fei-Fei Li 11 Feb '09

11 Feb '09

-------- Original Message -------- Dear all, Stuart Geman, James Manning Professor of Applied Mathematics from Brown University, will give a CS Colloquium talk at 4:15pm TODAY in the CS Small Auditorium. Prof. Geman has been a leader in statistics and computer vision. His recent projects span from computer vision to neuroscience to financial markets. Below please find his talk title and abstracts. Fei-Fei Google and the Vapnik-Chervonenkis Dimension Stuart Geman Brown University Google engineers routinely train query classifiers, for ranking advertisements or search results, on more words than any human being sees or hears in a lifetime. A human being who sees a meaningfully new image every second for one-hundred years will not see as many images as Google has in its libraries, all of which are available for training object detectors and image classifiers. Yet by human standards the state-of-the-art, in computer understanding of language and computer-generated image analysis, is primitive. What explains the gap? Why can’t learning theory tell us how to make machines that learn as efficiently as humans? Upper bounds on the number of training samples needed to learn a classifier as rich and competent as the human visual system can be derived using the Vapnik-Chervonenkis dimension, or the metric entropy, but these suggest that not only does Google need more examples, but all of evolution might fall short. I will make some proposals for efficient learning and offer some mathematics to support them. _______________________________________________ Ml-stat-talks mailing list Ml-stat-talks(a)lists.cs.princeton.edu https://lists.cs.princeton.edu/mailman/listinfo/ml-stat-talks -- --------------------------------------------- Li, Fei-Fei Ph.D. (publish under L. Fei-Fei) Assistant Professor Computer Science Dept. Princeton University 35 Olden St. Princeton, NJ 08540 Tel: (609)258-8130 Office: Room 411, CS Building Website: http://vision.cs.princeton.edu ---------------------------------------------

1 0

reminder: Tomorrow's CS Colloquium (Feb 11): Stuart Geman]
by Fei-Fei Li 10 Feb '09

10 Feb '09

-------- Original Message -------- Subject: [Ml-stat-talks] CS Colloquium (Feb 11): Stuart Geman Date: Sun, 01 Feb 2009 17:08:43 -0500 From: Fei-Fei Li <feifeili(a)CS.Princeton.EDU> CC: csugrad(a)CS.Princeton.EDU, ml-stat-talks(a)lists.cs.princeton.edu, csgrad(a)CS.Princeton.EDU, csfac(a)CS.Princeton.EDU, pixl-talks(a)lists.cs.princeton.edu References: <8697FFE6-7521-4BB6-A8E8-7C1A55376400(a)cs.princeton.edu> Dear all, Stuart Geman, James Manning Professor of Applied Mathematics from Brown University, will give a CS Colloquium talk at 4:15pm Wednesday Feb 11. Prof. Geman has been a leader in statistics and computer vision. His recent projects span from computer vision to neuroscience to financial markets. Below please find his talk title and abstracts. Fei-Fei Google and the Vapnik-Chervonenkis Dimension Stuart Geman Brown University Google engineers routinely train query classifiers, for ranking advertisements or search results, on more words than any human being sees or hears in a lifetime. A human being who sees a meaningfully new image every second for one-hundred years will not see as many images as Google has in its libraries, all of which are available for training object detectors and image classifiers. Yet by human standards the state-of-the-art, in computer understanding of language and computer-generated image analysis, is primitive. What explains the gap? Why can’t learning theory tell us how to make machines that learn as efficiently as humans? Upper bounds on the number of training samples needed to learn a classifier as rich and competent as the human visual system can be derived using the Vapnik-Chervonenkis dimension, or the metric entropy, but these suggest that not only does Google need more examples, but all of evolution might fall short. I will make some proposals for efficient learning and offer some mathematics to support them.

1 0

Speaker Wednesday - 2/11
by Ginny Hogan 09 Feb '09

09 Feb '09

Talk Wed. - Small Auditorium (Room 105) 4:15pm Host: Fei-Fei Li Stuart Geman, Brown University - "Google and the Vapnik-Chervonenkis Dimension"

1 0

Colloquium Speaker Today -
by Ginny Hogan 04 Feb '09

04 Feb '09

Andrew Ng, Stanford, will present his Talk "STAIR: The Stanford Artificial Intelligence Robot Project" at 4:15pm today COS 105 -

1 0