From macglashan at tti-c.org Mon Nov 3 09:22:42 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Nov 3 09:20:40 2008 Subject: [TTIC Colloquium] TTI-C Colloquium: David Forsyth, UIUC References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: When: TODAY: Monday, November 3 @ 2:00pm Where: TTI-C Conference Room: 1427 E. 60th St, 2nd Floor Who: David Forsyth, University of Illinois at Urbana-Champaign Title: Looking at People There is a great need for programs that can describe what people are doing from video. This is difficult to do, because it is hard to identify and track people in video sequences, because we have no canonical vocabulary for describing what people are doing, and because phenomena such as aspect and individual variation greatly affect the appearance of what people are doing. Recent work in kinematic tracking has produced methods that can report the kinematic configuration of the body fairly accurately and fully automatically. The problem of vocabulary is more difficult. I will discuss a generative activity model that allows activities to be assembled from a set of distinct spatial and temporal components. The models themselves are learned from labelled motion capture data and are assembled in a way that makes it possible to learn very complex finite automata without estimating large numbers of parameters. The advantage of such a model is that one can search videos for examples of activities specified with a simple query language, without possessing any example of the activity sought. In this case, aspect is dealt with by explicit 3D reasoning. An alternative strategy for dealing with aspect and individual variation is to build discriminative methods applied to appearance features. The difficulty here is that activities look different when seen from different directions. I will describe recent methods that make it possible to transfer models --- that is, to learn a model of an activity from one view, then recognize it in a completely different view. Contact: Greg Shakhnarovich, TTI-C greg@tti-c.org 834-2572 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20081103/ff4ec697/attachment-0001.htm From macglashan at tti-c.org Wed Nov 5 09:34:56 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Wed Nov 5 09:32:42 2008 Subject: [TTIC Colloquium] UC Talk: Ketan Mulmuley Message-ID: <1910EC83C1BF44CCB8998A5DCDA01E57@jmacglDPLFYD1> DEPARTMENT OF COMPUTER SCIENCE UNIVERSITY OF CHICAGO Date: TODAY, Wednesday, November 5, 2008 Time: 2:30 - 3:50 p.m. Place: RY 251 ---------------------------------------------------------- Speaker: Ketan Mulmuley From: University of Chicago Web page: http://www.cs.uchicago.edu/people/mulmuley Title: On P vs NP, Geometric Complexity theory, and the Riemann Hypothesis Abstract: This series of three colloquium talks on November 5, 12 and 19 (2.30 p.m.) will give a nontechnical, high level overview of geometric complexity theory (GCT), which is an approach to the P vs. NP problem via algebraic geometry, representation theory, and the theory of a new class of quantum groups, called nonstandard quantum groups, that arise in this approach. In particular, GCT says that the P vs. NP problem in characteristic zero is intimately linked to the Riemann Hypothesis over finite fields. A high level view of potential implications in mathematics, physics and quantum computation would also be given. No background in algebraic geometry, representation theory or quantum groups would be assumed. Complementary talks in the logic and theory seminars on November 10 (at 2.30 p.m. and 3.45 p.m.) would elaborate on the basic notion of obstructions in GCT. References for GCT: The basic plan of GCT is given in: GCTflip: "On P vs. NP, Geometric Complexity Theory and the Flip I: high level view". It has been partially implemented in a series of papers: GCT1 to GCT11. GCT1 to 4: Joint with Milind Sohoni GCT5: Joint with Hari Narayanan GCTflip, its abstract (GCTabs), and GCT1-8 are available on the speaker's personal home page. GCT8-11 are under preparation. -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20081105/1f3487fd/attachment.htm From macglashan at tti-c.org Wed Nov 5 09:39:21 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Wed Nov 5 09:37:06 2008 Subject: [TTIC Colloquium] ML Seminar: John Blitzer, UC Berkeley References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: When: Thursday, November 6 @ 2:00pm Where: TTI-C Conference Room: 1427 E. 60th St, 2nd Floor Who: John Blitzer, U.C. Berkeley Title: Adapting Natural Language Processing Systems to New Domains Statistical language processing tools are being applied to an ever-wider and more varied range of linguistic data. Researchers and engineers are using statistical models to organize and understand financial news, legal documents, biomedical abstracts, and weblog entries, among many other domains. Because language varies so widely, collecting and curating training sets for each different domain is prohibitively expensive. At the same time, differences in vocabulary and writing style across domains cause state-of-the-art supervised models to dramatically increase in error. The first part of this talk describes structural correspondence learning (SCL), a method for adapting linear discriminative models from resource-rich source domains to resource-poor target domains. The key idea is the use of pivot features which occur frequently and behave similarly in both the source and target domains. SCL builds a shared representation by searching for a low-dimensional feature subspace that allows us to accurately predict the presence or absence of pivot features on unlabeled data. We demonstrate SCL on the problem of sentiment classification for product reviews, and we show a more than 30% relative reduction in error for adapting models from one type of product to another. In the second part of the talk, we will describe a formal framework for analyzing domain adaptation tasks. We'll first introduce a measure of divergence that depends on the feature space from which we estimate our supervised model. Then we use this measure to state an upper bound on the true target error of a model trained to minimize a convex combination of empirical source and target errors. The bound characterizes the tradeoff inherent in training on both the large quantity of biased source data and the small quantity of unbiased target data, and we can compute it from finite labeled and unlabeled samples of the source and target distributions under relatively weak assumptions. Finally, we confirm experimentally that the bound corresponds well to empirical target error for the sentiment classification problem. Contact: Shai Shalev-Shwartz, TTI-C shai@tti-c.org 834-6850 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20081105/7d3c6a03/attachment.htm From macglashan at tti-c.org Thu Nov 6 08:55:02 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Thu Nov 6 08:52:50 2008 Subject: [TTIC Colloquium] TTI-C Colloquium: Ali Rahimi, Intel References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <9B158C3B2AA94711B1B7DE36F3D3903F@jmacglDPLFYD1> When: Monday, November 10 @ 2:00pm Where: TTI-C Conference Room: 1427 E. 60th St, 2nd Floor Who: Ali Rahimi, Intel Title: Random Features: Replacing Optimization with Randomization in Learning Training modern supervised learning models like weighted sums of kernels (as in the Kernelized SVM) and ensembles of weak learners (as in Adaboost) typically requires carrying out a meticulous optimization over a large number of parameters. But there is a much simpler way: instead of optimizing over all the parameters, I propose to randomize over most of parameters and then carry out a much cheaper optimization over the rest. A theoretical analysis for this Random Features trick using the concentration of measure phenomenon in Banach spaces guarantees that doing this trick almost as good as carrying out the full optimization. The empirical performance is even more surprising: on moderate-sized datasets (~60,000 examples), we get speedups of three orders of magnitude with no loss in accuracy, and we can train on datasets with millions of examples in a few minutes. I'll also briefly mention other machine learning and vision projects at Intel's Berkeley and Seattle lablets, including a real-time object recognition system, a large-scale 3D reconstruction of Seattle, data reduction tricks to speed up large scale clustering, theoretical guarantees for kernel machines when the kernel is not positive definite, and an analysis of the execution of faulty CPU using online learning bounds. Contact: Nati Srebro, TTI-C nati@tti-c.org 834-7493 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20081106/57697657/attachment-0001.htm From macglashan at tti-c.org Fri Nov 7 14:55:14 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Fri Nov 7 14:53:01 2008 Subject: [TTIC Colloquium] UC Seminar CANCELLATION: Ketan Mulmuley Message-ID: <113CFABE450E43489F18ADCE54526AE5@jmacglDPLFYD1> ***The remaining talks in this series, Nov 12th and 19th, have been cancelled*** _____ DEPARTMENT OF COMPUTER SCIENCE UNIVERSITY OF CHICAGO Date: Wednesday, November 5, 12 and 19, 2008 Time: 2:30 p.m. Place: RY 251 ---------------------------------------------------------- Speaker: Ketan Mulmuley From: University of Chicago Web page: http://www.cs.uchicago.edu/people/mulmuley Title: On P vs NP, Geometric Complexity theory, and the Riemann Hypothesis Abstract: This series of three colloquium talks on November 5, 12 and 19 (2.30 p.m.) will give a nontechnical, high level overview of geometric complexity theory (GCT), which is an approach to the P vs. NP problem via algebraic geometry, representation theory, and the theory of a new class of quantum groups, called nonstandard quantum groups, that arise in this approach. In particular, GCT says that the P vs. NP problem in characteristic zero is intimately linked to the Riemann Hypothesis over finite fields. A high level view of potential implications in mathematics, physics and quantum computation would also be given. No background in algebraic geometry, representation theory or quantum groups would be assumed. Complementary talks in the logic and theory seminars on November 10 (at 2.30 p.m. and 3.45 p.m.) would elaborate on the basic notion of obstructions in GCT. References for GCT: The basic plan of GCT is given in: GCTflip: "On P vs. NP, Geometric Complexity Theory and the Flip I: high level view". It has been partially implemented in a series of papers: GCT1 to GCT11. GCT1 to 4: Joint with Milind Sohoni GCT5: Joint with Hari Narayanan GCTflip, its abstract (GCTabs), and GCT1-8 are available on the speaker's personal home page. GCT8-11 are under preparation. -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20081107/aa9db6c2/attachment.htm From macglashan at tti-c.org Mon Nov 10 08:50:00 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Nov 10 08:47:32 2008 Subject: [TTIC Colloquium] TTI-C Colloquium: Ali Rahimi, Intel References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: When: TODAY: Monday, November 10 @ 2:00pm Where: TTI-C Conference Room: 1427 E. 60th St, 2nd Floor Who: Ali Rahimi, Intel Title: Random Features: Replacing Optimization with Randomization in Learning Training modern supervised learning models like weighted sums of kernels (as in the Kernelized SVM) and ensembles of weak learners (as in Adaboost) typically requires carrying out a meticulous optimization over a large number of parameters. But there is a much simpler way: instead of optimizing over all the parameters, I propose to randomize over most of parameters and then carry out a much cheaper optimization over the rest. A theoretical analysis for this Random Features trick using the concentration of measure phenomenon in Banach spaces guarantees that doing this trick almost as good as carrying out the full optimization. The empirical performance is even more surprising: on moderate-sized datasets (~60,000 examples), we get speedups of three orders of magnitude with no loss in accuracy, and we can train on datasets with millions of examples in a few minutes. I'll also briefly mention other machine learning and vision projects at Intel's Berkeley and Seattle lablets, including a real-time object recognition system, a large-scale 3D reconstruction of Seattle, data reduction tricks to speed up large scale clustering, theoretical guarantees for kernel machines when the kernel is not positive definite, and an analysis of the execution of faulty CPU using online learning bounds. Contact: Nati Srebro, TTI-C nati@tti-c.org 834-7493 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20081110/ea3e18fa/attachment.htm From macglashan at tti-c.org Mon Nov 10 15:43:55 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Nov 10 15:41:30 2008 Subject: [TTIC Colloquium] UC Seminar: Dr. Lee Rosen Message-ID: Seminar at the University of Chicago Peer Review at NIH Lee Rosen, Ph.D. Scientific Review Officer Biomedical Imaging Technology (BMIT) Surgical Sciences, Biomedical Imaging and Bioengineering (SBIB) NIH Center for Scientific Review Time: Dec 2nd, Tuesday at 11:00 am. Place: Radiology Resident Conference Room, (Q207), UC Medical Center 5812 South Ellis Avenue, Chicago, IL 60637 Dr. Lee Rosen, who is the Scientific Review Administrator for the NIH study section BMIT (Biomedical Imaging Technology) will be speaking at The University of Chicago on Dec 2nd, Tuesday at 11:00 am. The BMIT study section is where most NIH medical imaging technology grant applications are reviewed. As a side note, the MEDI (Medical Imaging) sister study-section is where NIH medical research grant applications that use imaging are reviewed. In the talk, Dr. Rosen will discuss the NIH peer review process and especially the anticipated, significant changes in NIH grant applications that will be implemented in the next several years. His presentation will include much information that will be of value particularly to junior investigators. After the presentation, Dr. Rosen has agreed to stay and meet with junior or other investigators who are planning to submit NIH grant applications. This is a great opportunity to meet the person you may be calling on the phone with questions about your grant or to ask questions about the review process specific to medical imaging. Junior faculty, post-doctoral fellows, and graduate students are particularly encouraged to attend. Biography: Dr. Lee Rosen serves as the Scientific Review Administrator for CSR's Biomedical Imaging Technology Study Section (BMIT). He also coordinates the review of electromagnetic device Small Business Innovative Research grants and technology resources. After earning his Ph.D. in physiology, Dr. Rosen had postdoctoral training in the Department of Pathology at Case Western Reserve University, with research in cardiovascular endothelial physiology. He then went into the private sector, where he worked primarily in small businesses, consulting with different Federal agencies. Since joining the Division of Research Grants in 1989, now the Center for Scientific Review, he has focused on imaging technology, with involvement in almost all types of grant mechanisms reviewed by CSR. -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20081110/40b73412/attachment-0001.htm From macglashan at tti-c.org Mon Nov 10 15:45:54 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Nov 10 15:43:19 2008 Subject: [TTIC Colloquium] Reformatted: UC Seminar, Lee Rosen Message-ID: <914D2B2BF4154C15869952115BFE290B@jmacglDPLFYD1> Seminar at The University of Chicago Peer Review at NIH Lee Rosen, Ph.D. Scientific Review Officer Biomedical Imaging Technology (BMIT) Surgical Sciences, Biomedical Imaging and Bioengineering (SBIB) NIH Center for Scientific Review Time: Dec 2nd, Tuesday at 11:00 am. Place: Radiology Resident Conference Room, (Q207), UC Medical Center 5812 South Ellis Avenue, Chicago, IL 60637 Dr. Lee Rosen, who is the Scientific Review Administrator for the NIH study section BMIT (Biomedical Imaging Technology) will be speaking at The University of Chicago on Dec 2nd, Tuesday at 11:00 am. The BMIT study section is where most NIH medical imaging technology grant applications are reviewed. As a side note, the MEDI (Medical Imaging) sister study-section is where NIH medical research grant applications that use imaging are reviewed. In the talk, Dr. Rosen will discuss the NIH peer review process and especially the anticipated, significant changes in NIH grant applications that will be implemented in the next several years. His presentation will include much information that will be of value particularly to junior investigators. After the presentation, Dr. Rosen has agreed to stay and meet with junior or other investigators who are planning to submit NIH grant applications. This is a great opportunity to meet the person you may be calling on the phone with questions about your grant or to ask questions about the review process specific to medical imaging. Junior faculty, post-doctoral fellows, and graduate students are particularly encouraged to attend. Biography: Dr. Lee Rosen serves as the Scientific Review Administrator for CSR's Biomedical Imaging Technology Study Section (BMIT). He also coordinates the review of electromagnetic device Small Business Innovative Research grants and technology resources. After earning his Ph.D. in physiology, Dr. Rosen had postdoctoral training in the Department of Pathology at Case Western Reserve University, with research in cardiovascular endothelial physiology. He then went into the private sector, where he worked primarily in small businesses, consulting with different Federal agencies. Since joining the Division of Research Grants in 1989, now the Center for Scientific Review, he has focused on imaging technology, with involvement in almost all types of grant mechanisms reviewed by CSR. From macglashan at tti-c.org Tue Nov 11 13:34:23 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Nov 11 13:32:19 2008 Subject: [TTIC Colloquium] ML Seminar Tomorrow: Shai Shalev-Shwartz, TTI-C References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <7805C8249B894E74AB987F6BBF9F0014@jmacglDPLFYD1> When: Tomorrow, Wednesday, Nov 12, 11:00am Where: TTI-C Conference Room: 1427 E. 60th St, 2nd Floor Who: Shai Shalev-Shwartz, TTI-C Title: Trading Accuracy for Sparsity Although many features might be available for use in a prediction task, it is often beneficial to use only a small subset of the available features. Predictors that use only a small subset of features require a smaller memory footprint and can be applied faster. Furthermore, in applications such as medical diagnostics, obtaining each possible ``feature'' (e.g. test result) can be costly, and so a predictor that uses only a small number of features is desirable, even at the cost of a small degradation in performance relative to a predictor that uses more features. In this work, we describe and analyze efficient methods for maximizing the accuracy of a predictor subject to a constraint on the number of features allowed. We provide an analysis of the tradeoff between accuracy and sparsity under several sets of assumptions on the data distribution and loss function. We contrast our results with (quite different) results regarding the sparsistency of the Lasso and compressed sensing. Joint work with Nati Srebro and Tong Zhang Contact: Shai Shalev-Shwartz, TTI-C shai@tti-c.org 834-6850 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20081111/116dfe6d/attachment.htm From macglashan at tti-c.org Tue Nov 18 09:12:24 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Nov 18 09:12:20 2008 Subject: [TTIC Colloquium] TTI-C Colloquium: Michael Mahoney, Stanford University References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <9FE074C851FD40918BE1420A8A074DF0@jmacglDPLFYD1> When: Monday, November 24 @ 2:00pm Where: TTI-C Conference Room: 1427 E. 60th St, 2nd Floor Who: Michael Mahoney, Stanford University Title: COMMUNITY STRUCTURE IN LARGE SOCIAL AND INFORMATION NETWORKS The concept of a community is central to social network analysis, and thus a large body of work has been devoted to identifying community structure. For example, a community may be thought of as a set of web pages on related topics, a set of people who share common interests, or more generally as a set of nodes in a network more similar amongst themselves than with the remainder of the network. Motivated by difficulties we experienced at actually finding meaningful communities in large real-world networks, we have performed a large scale analysis of a wide range of social and information networks. Our main methodology uses local spectral methods and involves computing isoperimetric properties of the networks at various size scales -- a novel application of ideas from scientific computation to internet data analysis. Our empirical results suggest a significantly more refined picture of community structure than has been appreciated previously. Our most striking finding is that in nearly every network dataset we examined, we observe tight but almost trivial communities at very small size scales, and at larger size scales, the best possible communities gradually ``blend in'' with the rest of the network and thus become less ``community-like.'' This behavior is not explained, even at a qualitative level, by any of the commonly-used network generation models. Moreover, this behavior is exactly the opposite of what one would expect based on experience with and intuition from expander graphs, from graphs that are well-embeddable in a low-dimensional structure, and from small social networks that have served as testbeds of community detection algorithms. Possible mechanisms for reproducing our empirical observations will be discussed, as will implications of these findings for clustering, classification, and more general data analysis in modern large social and information networks. Contact: Nati Srebro, TTI-C nati@tti-c.org 834-7493 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20081118/ba7f645e/attachment.htm From macglashan at tti-c.org Tue Nov 18 10:24:17 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Nov 18 10:24:17 2008 Subject: [TTIC Colloquium] ML Seminar: Steve Smale Message-ID: <99F5B65A54D04E3BBDA0F750CD3D8775@jmacglDPLFYD1> When: Tomorrow- Wed, Nov 19 @ 11:00am Where: TTI-C conference room Speaker: Steve Smale Title: Understanding Patterns in Data Abstract: We will show how patterns in data can be better understood using topological and geometrical methods (although we are not assuming any prior knowledge of topology or geometry). From macglashan at tti-c.org Mon Nov 24 09:11:51 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Nov 24 09:11:29 2008 Subject: [TTIC Colloquium] TTI-C Colloquium: Michael Mahoney, Stanford University References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <79C3CFB6C01B445ABC93CC1C61D54BDB@jmacglDPLFYD1> When: TODAY: Monday, November 24 @ 2:00pm Where: TTI-C Conference Room: 1427 E. 60th St, 2nd Floor Who: Michael Mahoney, Stanford University Title: COMMUNITY STRUCTURE IN LARGE SOCIAL AND INFORMATION NETWORKS The concept of a community is central to social network analysis, and thus a large body of work has been devoted to identifying community structure. For example, a community may be thought of as a set of web pages on related topics, a set of people who share common interests, or more generally as a set of nodes in a network more similar amongst themselves than with the remainder of the network. Motivated by difficulties we experienced at actually finding meaningful communities in large real-world networks, we have performed a large scale analysis of a wide range of social and information networks. Our main methodology uses local spectral methods and involves computing isoperimetric properties of the networks at various size scales -- a novel application of ideas from scientific computation to internet data analysis. Our empirical results suggest a significantly more refined picture of community structure than has been appreciated previously. Our most striking finding is that in nearly every network dataset we examined, we observe tight but almost trivial communities at very small size scales, and at larger size scales, the best possible communities gradually ``blend in'' with the rest of the network and thus become less ``community-like.'' This behavior is not explained, even at a qualitative level, by any of the commonly-used network generation models. Moreover, this behavior is exactly the opposite of what one would expect based on experience with and intuition from expander graphs, from graphs that are well-embeddable in a low-dimensional structure, and from small social networks that have served as testbeds of community detection algorithms. Possible mechanisms for reproducing our empirical observations will be discussed, as will implications of these findings for clustering, classification, and more general data analysis in modern large social and information networks. Contact: Nati Srebro, TTI-C nati@tti-c.org 834-7493 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20081124/0156c0d7/attachment-0001.htm