From macglashan at tti-c.org Thu May 1 09:40:22 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Thu May 1 10:39:01 2008 Subject: [TTIC Colloquium] Vahab Mirrokni (Theory Group, Microsoft Research)- TTI-C Talk Message-ID: <000e01c8ab99$4babdec0$aabf8780@jmacglDPLFYD1> When: ???Thursday, May 8 @ 10:00am Where:???TTI-C Conference Room, 1427 E. 60th St. Who:?????Vahab Mirrokni (Theory Group, Microsoft Research) Topic:? ?Online Advertisement and Submodular Maximization Submodular maximization is a central problem in optimization with many applications in data mining, social network analysis, and online advertisement. Unlike the problem of minimizing submodular functions, the problem of maximizing submodular functions is NP-hard. We design the first constant-factor approximation algorithms for maximizing Non-negative submodular functions. In particular, we give a deterministic local search 1/3-approximation and a randomized 2/5-approximation algorithm for maximizing non-negative submodular functions. Furthermore, we prove that achieving an approximation factor better than 1/2 requires exponential time. Then, I will discuss applications of submodular maximization in the growing field of the online advertisement, and in particular two specific applications in marketing digital goods over social networks, and revenue maximization for guaranteed banner advertisement. The first application is concerned with viral marketing and word-of-mouth advertising in social networks. The second application is related to the banner ad allocation problem satisfying a guaranteed delivery property. The main part of the talk is based on joint work with Feige and Vondrak (FOCS 2007), Hartline and Sundararajan (WWW 2008), and Feige, Immorlica, and Nazerzadeh (WWW 2008). Bio: Vahab Mirrokni is a Postdoctoral Researcher in the Theory Group at Microsoft Research. He received his PhD from Massachusetts Institute of Technology and his B.Sc. from Sharif University of Technology. During his PhD studies, he spent some time doing research at IBM Research, Bell-Laboratories, Microsoft Research, and Amazon.com. His research areas include algorithmic game theory, approximation algorithms, and social network analysis. Recently at Microsoft Research, he has been working on various algorithmic and economic problems related to the Internet search and online advertisement. He has published over 50 papers and has filed more than 10 patents. Contact:?????Julia Chuzhoy, TTI-C?????cjulia@tti-c.org?????834-2490 From macglashan at tti-c.org Wed May 7 10:33:43 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Wed May 7 11:31:57 2008 Subject: [TTIC Colloquium] Vahab Mirrokni (Theory Group, Microsoft Research)- TTI-C Talk Message-ID: <002901c8b057$bbf63fa0$aabf8780@jmacglDPLFYD1> When: ???Thursday, May 8 @ 10:00am Where:???TTI-C Conference Room, 1427 E. 60th St. Who:?????Vahab Mirrokni (Theory Group, Microsoft Research) Topic:? ?Online Advertisement and Submodular Maximization Submodular maximization is a central problem in optimization with many applications in data mining, social network analysis, and online advertisement. Unlike the problem of minimizing submodular functions, the problem of maximizing submodular functions is NP-hard. We design the first constant-factor approximation algorithms for maximizing Non-negative submodular functions. In particular, we give a deterministic local search 1/3-approximation and a randomized 2/5-approximation algorithm for maximizing non-negative submodular functions. Furthermore, we prove that achieving an approximation factor better than 1/2 requires exponential time. Then, I will discuss applications of submodular maximization in the growing field of the online advertisement, and in particular two specific applications in marketing digital goods over social networks, and revenue maximization for guaranteed banner advertisement. The first application is concerned with viral marketing and word-of-mouth advertising in social networks. The second application is related to the banner ad allocation problem satisfying a guaranteed delivery property. The main part of the talk is based on joint work with Feige and Vondrak (FOCS 2007), Hartline and Sundararajan (WWW 2008), and Feige, Immorlica, and Nazerzadeh (WWW 2008). Bio: Vahab Mirrokni is a Postdoctoral Researcher in the Theory Group at Microsoft Research. He received his PhD from Massachusetts Institute of Technology and his B.Sc. from Sharif University of Technology. During his PhD studies, he spent some time doing research at IBM Research, Bell-Laboratories, Microsoft Research, and Amazon.com. His research areas include algorithmic game theory, approximation algorithms, and social network analysis. Recently at Microsoft Research, he has been working on various algorithmic and economic problems related to the Internet search and online advertisement. He has published over 50 papers and has filed more than 10 patents. Contact:?????Julia Chuzhoy, TTI-C?????cjulia@tti-c.org?????834-2490 From macglashan at tti-c.org Thu May 8 13:57:18 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Thu May 8 14:55:26 2008 Subject: [TTIC Colloquium] Tom Hayes (TTI-C) Talk Message-ID: <002f01c8b13d$57340ed0$aabf8780@jmacglDPLFYD1> When: ???Friday, May 9 @ 10:00am Where:???TTI-C Conference Room, 1427 E. 60th St. Who:?????Tom Hayes, TTI-C Topic:? ?The Forgiving Tree: A Distributed Algorithm for Self-Healing in Reconfigurable Networks We will study an abstract model for cascading node failures in enormous graphs, motivated by examples such as the Skype network failure for two days last August, or the California "electricity crisis" of 2000-2001. In this model, our algorithm will repeatedly respond to adversarial node deletions by quickly adding a few new edges with the following goals: (1) maintain connectivity of the network (2) avoid large increases in node degrees (which might cause further failures) (3) prevent large increases in the diameter of the network. Here, by "quickly," (4) we allow the response time to depend on the degree of the deleted node, but not on the size of the network. I will describe a new algorithm, called the Forgiving Tree, which achieves these goals. Over the course of any sequence of node deletions: (1) the network stays connected, (2) no node degree increases by more than a total of 3, (3) the diameter doesn't increase by more than log(max degree), and (4) the responses require O(1) messages of size O(1) to be sent and received by each neighbor of the deleted node (these nodes and their neighbors are the only ones involved in the self-healing step). Our result has several advantages over more traditional approaches. (1) rather than *designing* a network which is robust to node failures, which typically requires a lot more edges, in our setting, we start out with any network topology, and preserve it as well as we can. (2) since there is no centralized control of the healing process, it can proceed much faster. (3) our bounds apply to adversarial sequences of node failures, rather than, e.g., random node failures. Joint work with Navin Rustagi, Jared Saia and Amitabh Trehan Contact:?????Tom Hayes, TTI-C?????hayest@tti-c.org?????834-3485 From macglashan at tti-c.org Mon May 19 09:02:03 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon May 19 09:59:44 2008 Subject: [TTIC Colloquium] Aryeh Kontorovich (Weizmann Institute): TTI-C Talk Message-ID: <000b01c8b9b8$ea8d50f0$aabf8780@jmacglDPLFYD1> When: Thursday, May 22 @ 11:00am Where: TTI-C Conference Room, 1427 E. 60th St., 2nd Floor Who: Aryeh Kontorovich, Weizmann Institute of Science Topic: A Universal Kernel for Learning Regular Languages We develop a novel approach to the classical problem of learning regular languages from labeled samples. Rather than attempting to construct small consistent automata (long known to be a very difficult computational problem), we embed the strings in a Hilbert space and compute a maximum-margin hyperplane, which becomes our classifier for new strings. We accomplish this via a universal kernel that renders all regular languages linearly separable. Under this kernel, the image of every regular language is linearly separable from its complement in some finite-dimensional space with a strictly positive margin. Thus, we are able to efficiently (in sample size) compute a maximum-margin separating hyperplane (via SVM, for example) and use margin bounds to control the generalization error. A brute-force computation of this universal kernel has super-exponential complexity. We conjecture that this problem is intractable (a likely candidate for #P-complete). However, we propose a simple randomized scheme for efficiently obtaining an $\eps$-approximation to our universal kernel. We show that the approximate kernel preserves the distances and margins with low distortion, and therefore may be used as a surrogate for the original one. To our knowledge, the framework we propose is the only one capable of inducing unrestricted regular languages from labeled samples (modulo standard cryptographic limitations). Along the way, we touch upon several fundamental questions in complexity, automata, and machine learning. Join work with Boaz Nadler. 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/20080519/b3a18e10/attachment.htm From macglashan at tti-c.org Wed May 21 09:35:43 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Wed May 21 10:33:14 2008 Subject: [TTIC Colloquium] Aryeh Kontorovich (Weizmann Institute): TTI-C Talk Message-ID: <000701c8bb4f$f372b2d0$aabf8780@jmacglDPLFYD1> When: Thursday, May 22 @ 11:00am Where: TTI-C Conference Room, 1427 E. 60th St., 2nd Floor Who: Aryeh Kontorovich, Weizmann Institute of Science Topic: A Universal Kernel for Learning Regular Languages We develop a novel approach to the classical problem of learning regular languages from labeled samples. Rather than attempting to construct small consistent automata (long known to be a very difficult computational problem), we embed the strings in a Hilbert space and compute a maximum-margin hyperplane, which becomes our classifier for new strings. We accomplish this via a universal kernel that renders all regular languages linearly separable. Under this kernel, the image of every regular language is linearly separable from its complement in some finite-dimensional space with a strictly positive margin. Thus, we are able to efficiently (in sample size) compute a maximum-margin separating hyperplane (via SVM, for example) and use margin bounds to control the generalization error. A brute-force computation of this universal kernel has super-exponential complexity. We conjecture that this problem is intractable (a likely candidate for #P-complete). However, we propose a simple randomized scheme for efficiently obtaining an $\eps$-approximation to our universal kernel. We show that the approximate kernel preserves the distances and margins with low distortion, and therefore may be used as a surrogate for the original one. To our knowledge, the framework we propose is the only one capable of inducing unrestricted regular languages from labeled samples (modulo standard cryptographic limitations). Along the way, we touch upon several fundamental questions in complexity, automata, and machine learning. Join work with Boaz Nadler. 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/20080521/849030e8/attachment.htm From macglashan at tti-c.org Wed May 21 12:00:33 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Wed May 21 12:58:03 2008 Subject: [TTIC Colloquium] DATE CHANGE! Aryeh Kontorovich (Weizmann Institute), TTI-C Talk Message-ID: <001201c8bb64$2f2cd490$aabf8780@jmacglDPLFYD1> ***DATE CHANGE: This talk has been rescheduled to next Wednesday May 28 @11:00am.*** _____ When: Thursday, May 22 @ 11:00am Rescheduled to Wednesday, May 28 @11:00am Where: TTI-C Conference Room, 1427 E. 60th St., 2nd Floor Who: Aryeh Kontorovich, Weizmann Institute of Science Topic: A Universal Kernel for Learning Regular Languages We develop a novel approach to the classical problem of learning regular languages from labeled samples. Rather than attempting to construct small consistent automata (long known to be a very difficult computational problem), we embed the strings in a Hilbert space and compute a maximum-margin hyperplane, which becomes our classifier for new strings. We accomplish this via a universal kernel that renders all regular languages linearly separable. Under this kernel, the image of every regular language is linearly separable from its complement in some finite-dimensional space with a strictly positive margin. Thus, we are able to efficiently (in sample size) compute a maximum-margin separating hyperplane (via SVM, for example) and use margin bounds to control the generalization error. A brute-force computation of this universal kernel has super-exponential complexity. We conjecture that this problem is intractable (a likely candidate for #P-complete). However, we propose a simple randomized scheme for efficiently obtaining an $\eps$-approximation to our universal kernel. We show that the approximate kernel preserves the distances and margins with low distortion, and therefore may be used as a surrogate for the original one. To our knowledge, the framework we propose is the only one capable of inducing unrestricted regular languages from labeled samples (modulo standard cryptographic limitations). Along the way, we touch upon several fundamental questions in complexity, automata, and machine learning. Join work with Boaz Nadler. 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/20080521/3b0088f6/attachment-0001.htm From macglashan at tti-c.org Tue May 27 09:32:43 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue May 27 10:29:50 2008 Subject: [TTIC Colloquium] Alan Yuille (UCLA Stat Dept): TTI-C Talk Message-ID: <001401c8c006$8681f620$aabf8780@jmacglDPLFYD1> When: Thursday, May 29 @ 3:00pm Where: TTI-C Conference Room, 1427 E. 60th St., 2nd Floor Who: Alan Yuille, UCLA Statistics Department Topic: Discriminative Learning for Deformable Object Parsing Work with (Leo) Log Zhu. We propose hierarchical models for representing shape and appearance of deformable objects. These models are learnt in a discriminative manner using techniques such as structure-perceptron and max-margin. We describe rapid inference algorithms which can parse these models efficiently. These approaches are demonstrated on horses, humans, faces and a range of other objects. We evaluate the effectiveness of these models on standard benchmarked databases for tasks such as detection, segmentation, matching and parsing. Contact: David McAllester, TTI-C mcallester@tti-c.org 702-5562 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20080527/0a2a0bb1/attachment.htm From macglashan at tti-c.org Tue May 27 10:00:31 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue May 27 10:57:39 2008 Subject: [TTIC Colloquium] Aryeh Kontorovich (Weizmann Institute), TTI-C Talk Message-ID: <002701c8c00a$69589c80$aabf8780@jmacglDPLFYD1> ***This talk has been rescheduled to Wednesday May 28 @11:00am.*** _____ When: Wednesday, May 28 @11:00am Where: TTI-C Conference Room, 1427 E. 60th St., 2nd Floor Who: Aryeh Kontorovich, Weizmann Institute of Science Topic: A Universal Kernel for Learning Regular Languages We develop a novel approach to the classical problem of learning regular languages from labeled samples. Rather than attempting to construct small consistent automata (long known to be a very difficult computational problem), we embed the strings in a Hilbert space and compute a maximum-margin hyperplane, which becomes our classifier for new strings. We accomplish this via a universal kernel that renders all regular languages linearly separable. Under this kernel, the image of every regular language is linearly separable from its complement in some finite-dimensional space with a strictly positive margin. Thus, we are able to efficiently (in sample size) compute a maximum-margin separating hyperplane (via SVM, for example) and use margin bounds to control the generalization error. A brute-force computation of this universal kernel has super-exponential complexity. We conjecture that this problem is intractable (a likely candidate for #P-complete). However, we propose a simple randomized scheme for efficiently obtaining an $\eps$-approximation to our universal kernel. We show that the approximate kernel preserves the distances and margins with low distortion, and therefore may be used as a surrogate for the original one. To our knowledge, the framework we propose is the only one capable of inducing unrestricted regular languages from labeled samples (modulo standard cryptographic limitations). Along the way, we touch upon several fundamental questions in complexity, automata, and machine learning. Join work with Boaz Nadler. 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/20080527/15869667/attachment.htm From macglashan at tti-c.org Wed May 28 09:16:10 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Wed May 28 10:13:16 2008 Subject: [TTIC Colloquium] Alan Yuille (UCLA Stat Dept): TTI-C Talk Message-ID: <000101c8c0cd$616b4db0$aabf8780@jmacglDPLFYD1> When: Thursday, May 29 @ 3:00pm Where: TTI-C Conference Room, 1427 E. 60th St., 2nd Floor Who: Alan Yuille, UCLA Statistics Department Topic: Discriminative Learning for Deformable Object Parsing Work with (Leo) Log Zhu. We propose hierarchical models for representing shape and appearance of deformable objects. These models are learnt in a discriminative manner using techniques such as structure-perceptron and max-margin. We describe rapid inference algorithms which can parse these models efficiently. These approaches are demonstrated on horses, humans, faces and a range of other objects. We evaluate the effectiveness of these models on standard benchmarked databases for tasks such as detection, segmentation, matching and parsing. Contact: David McAllester, TTI-C mcallester@tti-c.org 702-5562 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20080528/dcafdd26/attachment-0001.htm