From macglashan at tti-c.org Tue Apr 1 10:42:02 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Apr 1 10:41:01 2008 Subject: [TTIC Colloquium] Narayana Santhanam, UC Berkeley- TTI-C Talk Message-ID: <001001c89417$50d368d0$aabf8780@jmacglDPLFYD1> When: Monday, April 7, 10:00am Where: TTI-C Conference Room Who: Narayana Santhanam, University of California, Berkeley Topic: High dimension statistical problems: practice and theory For advances in biology, computation and storage, we have invited the "curse of dimensionality" upon many problems that concern the modern engineer. The colorful phrase in quotes coined by Bellman refers to the usual inability of classical methods to handle problem instances wherein the number of parameters associated with each data sample is comparable to number of samples we have to work on. In this talk we focus on the problem of discrete distribution estimation in the undersampled regime, and develop theory to tackle this problem using ideas from information theory, number theory, combinatorics, analysis as well as tools in statistical learning. This framework encompasses well known algorithms including the Laplace and Good Turing estimator. We apply these approaches to classifying text, and obtain very fast algorithms that stand up to (and in many cases, beat) support vector machines in both performance and speed. The big picture is to see this work as source coding driven by data analysis, complementing the traditional communication/storage driven models. We conclude with a brief preview of some of the directions in which we are developing this work. Bio: Narayana Santhanam is a postdoctoral researcher hosted by Prof. Martin Wainwright in UC Berkeley. He obtained the B.Tech degree from IIT Madras, and MS and PhD with Prof. Alon Orlitsky from UC San Diego. He is interested in theory and applications related to high dimensional problems, statistical learning, information theory and combinatorial/probabilistic problems in general. He is the recipient of the 2006 Information theory society award and the 2003 Capocelli Prize. Contact: Nathan Srebro, TTI-C nati@tti-c.org 834-7493 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20080401/aa33bb09/attachment-0001.htm From macglashan at tti-c.org Tue Apr 1 14:45:35 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Apr 1 14:44:28 2008 Subject: [TTIC Colloquium] Andreas Krause, Carnegie Mellon University- TTI-C Talk Message-ID: <005301c89439$56570790$aabf8780@jmacglDPLFYD1> When: Tuesday, April 8, 1:00pm Where: TTI-C Conference Room Who: Andreas Krause, Carnegie Mellon University Topic: Optimizing Sensing from Water to the Web Where should we place sensors to quickly detect contaminations in drinking water distribution networks? Which blogs should we read to learn about the biggest stories on the web? These problems share a fundamental challenge: How can we obtain the most useful information about the state of the world, at minimum cost? Such sensing, or active learning, problems are typically NP-hard, and were commonly addressed using heuristics without theoretical guarantees about the solution quality. In this talk, I will present algorithms which efficiently find provably near-optimal solutions to large, complex sensing problems. Our algorithms exploit submodularity, an intuitive notion of diminishing returns, common to many sensing problems; the more sensors we have already deployed, the less we learn by placing another sensor. To quantify the uncertainty in our predictions, we use probabilistic models, such as Gaussian Processes. In addition to identifying the most informative sensing locations, our algorithms can handle more challenging settings, where sensors need to be able to reliably communicate over lossy links, where mobile robots are used for collecting data or where solutions need to be robust against adversaries and sensor failures. I will also present results applying our algorithms to several real-world sensing tasks, including environmental monitoring using robotic sensors, activity recognition using a built sensing chair, deciding which blogs to read on the web, and a sensor placement competition. Bio: Andreas Krause is a Ph.D. Candidate at the Computer Science Department of Carnegie Mellon University. He is a recipient of a Microsoft Research Graduate Fellowship, and his research on sensor placement and information acquisition received awards at several conferences (KDD '07, IPSN '06, ICML '05 and UAI '05). He obtained his Diplom in Computer Science and Mathematics from the Technische Universit?t M?nchen, where his research received the NRW Undergraduate Science Award. Contact: Nathan Srebro, TTI-C nati@tti-c.org 834-7493 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20080401/0c0baaa8/attachment.htm From macglashan at tti-c.org Thu Apr 3 09:15:41 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Thu Apr 3 10:15:57 2008 Subject: [TTIC Colloquium] Richard Waltz, University of Southern California- TTI-C Talk Message-ID: <000c01c8959d$953a5010$aabf8780@jmacglDPLFYD1> When: Thursday, April 10, 1:00pm Where: TTI-C Conference Room Who: Richard Waltz, University of Southern California Topic: New Active-Set Algorithms for Large-Scale Nonlinear Optimization Active-set algorithms offer a powerful approach for solving nonlinear optimization problems. These methods have many advantages over the more recently popular interior-point methods; most notably the ability to converge quickly (i.e., "warm start") from an advanced initial point. However, current active-set methods are unable to scale to large problem sizes as effectively as interior-point methods, and this significantly limits their applicability. This talk will present new techniques for identifying which inequality constraints are "active" (i.e., hold as equalities) at the solution of nonlinear optimization problems. These techniques, based on solving linear programming subproblems, allow the active-set estimate to change by many constraints at once and overcome the bottlenecks of traditional active-set methods. We also present advances in penalty methods used to relax constraints in nonlinear optimization models. These penalty methods are integrated with our new active-set identification techniques to form a novel active-set algorithm that outperforms traditional active-set methods on large-scale nonlinear optimization problems. Contact: Nathan Srebro, TTI-C nati@tti-c.org 834-7493 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20080403/522d264c/attachment.htm From macglashan at tti-c.org Fri Apr 4 10:26:17 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Fri Apr 4 11:26:24 2008 Subject: [TTIC Colloquium] Narayana Santhanam, UC Berkeley- TTI-C Talk Message-ID: <000d01c89670$9c3f5940$aabf8780@jmacglDPLFYD1> When: Monday, April 7, 10:00am Where: TTI-C Conference Room Who: Narayana Santhanam, University of California, Berkeley Topic: High dimension statistical problems: practice and theory For advances in biology, computation and storage, we have invited the "curse of dimensionality" upon many problems that concern the modern engineer. The colorful phrase in quotes coined by Bellman refers to the usual inability of classical methods to handle problem instances wherein the number of parameters associated with each data sample is comparable to number of samples we have to work on. In this talk we focus on the problem of discrete distribution estimation in the undersampled regime, and develop theory to tackle this problem using ideas from information theory, number theory, combinatorics, analysis as well as tools in statistical learning. This framework encompasses well known algorithms including the Laplace and Good Turing estimator. We apply these approaches to classifying text, and obtain very fast algorithms that stand up to (and in many cases, beat) support vector machines in both performance and speed. The big picture is to see this work as source coding driven by data analysis, complementing the traditional communication/storage driven models. We conclude with a brief preview of some of the directions in which we are developing this work. Bio: Narayana Santhanam is a postdoctoral researcher hosted by Prof. Martin Wainwright in UC Berkeley. He obtained the B.Tech degree from IIT Madras, and MS and PhD with Prof. Alon Orlitsky from UC San Diego. He is interested in theory and applications related to high dimensional problems, statistical learning, information theory and combinatorial/probabilistic problems in general. He is the recipient of the 2006 Information theory society award and the 2003 Capocelli Prize. Contact: Nathan Srebro, TTI-C nati@tti-c.org 834-7493 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20080404/401ff27e/attachment-0001.htm From cnovak at tti-c.org Mon Apr 7 09:48:10 2008 From: cnovak at tti-c.org (Christina Novak) Date: Mon Apr 7 10:48:12 2008 Subject: [TTIC Colloquium] Andreas Krause, Carnegie Mellon University- TTI-C Talk Message-ID: <004201c898be$66d82a70$a9bf8780@cnovakHBRQFD1> When: Tuesday, April 8, 1:00pm Where: TTI-C Conference Room Who: Andreas Krause, Carnegie Mellon University Topic: Optimizing Sensing from Water to the Web Where should we place sensors to quickly detect contaminations in drinking water distribution networks? Which blogs should we read to learn about the biggest stories on the web? These problems share a fundamental challenge: How can we obtain the most useful information about the state of the world, at minimum cost? Such sensing, or active learning, problems are typically NP-hard, and were commonly addressed using heuristics without theoretical guarantees about the solution quality. In this talk, I will present algorithms which efficiently find provably near-optimal solutions to large, complex sensing problems. Our algorithms exploit submodularity, an intuitive notion of diminishing returns, common to many sensing problems; the more sensors we have already deployed, the less we learn by placing another sensor. To quantify the uncertainty in our predictions, we use probabilistic models, such as Gaussian Processes. In addition to identifying the most informative sensing locations, our algorithms can handle more challenging settings, where sensors need to be able to reliably communicate over lossy links, where mobile robots are used for collecting data or where solutions need to be robust against adversaries and sensor failures. I will also present results applying our algorithms to several real-world sensing tasks, including environmental monitoring using robotic sensors, activity recognition using a built sensing chair, deciding which blogs to read on the web, and a sensor placement competition. Bio: Andreas Krause is a Ph.D. Candidate at the Computer Science Department of Carnegie Mellon University. He is a recipient of a Microsoft Research Graduate Fellowship, and his research on sensor placement and information acquisition received awards at several conferences (KDD '07, IPSN '06, ICML '05 and UAI '05). He obtained his Diplom in Computer Science and Mathematics from the Technische Universit?t M?nchen, where his research received the NRW Undergraduate Science Award. Contact: Nathan Srebro, TTI-C nati@tti-c.org 834-7493 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20080407/ef90295b/attachment.htm From nati at uchicago.edu Mon Apr 7 17:02:57 2008 From: nati at uchicago.edu (Nathan Srebro) Date: Mon Apr 7 18:02:47 2008 Subject: [TTIC Colloquium] **TIME CHANGE** Andreas Krause, Carnegie Mellon University- TTI-C Talk Message-ID: <154084920804071502g5905a514ka99ddd0a45110216@mail.gmail.com> Please note: to accommodate those who would like to attend also Eirc Vigoda's talk, Andreas Krause's talk will be postponed by 15 minutes. The talk will start at 1:15PM rather than 1PM. > > > > > > > When: Tuesday, April 8, 1:00pm > > > > Where: TTI-C Conference Room > > > > Who: Andreas Krause, Carnegie Mellon University > > > > Topic: Optimizing Sensing from Water to the Web > > > > Where should we place sensors to quickly detect contaminations in drinking > water distribution networks? Which blogs should we read to learn about the > biggest stories on the web? These problems share a fundamental challenge: > How can we obtain the most useful information about the state of the world, > at minimum cost? > > > > Such sensing, or active learning, problems are typically NP-hard, and were > commonly addressed using heuristics without theoretical guarantees about the > solution quality. In this talk, I will present algorithms which efficiently > find provably near-optimal solutions to large, complex sensing problems. Our > algorithms exploit submodularity, an intuitive notion of diminishing > returns, common to many sensing problems; the more sensors we have already > deployed, the less we learn by placing another sensor. To quantify the > uncertainty in our predictions, we use probabilistic models, such as > Gaussian Processes. > > > > In addition to identifying the most informative sensing locations, our > algorithms can handle more challenging settings, where sensors need to be > able to reliably communicate over lossy links, where mobile robots are used > for collecting data or where solutions need to be robust against adversaries > and sensor failures. > > > > I will also present results applying our algorithms to several real-world > sensing tasks, including environmental monitoring using robotic sensors, > activity recognition using a built sensing chair, deciding which blogs to > read on the web, and a sensor placement competition. > > > > > > > > Bio: > > > > Andreas Krause is a Ph.D. Candidate at the Computer Science Department of > Carnegie Mellon University. He is a recipient of a Microsoft Research > Graduate Fellowship, and his research on sensor placement and information > acquisition received awards at several conferences (KDD '07, IPSN '06, ICML > '05 and UAI '05). He obtained his Diplom in Computer Science and Mathematics > from the Technische Universit?t M?nchen, where his research received the NRW > Undergraduate Science Award. > > > > > > > > Contact: Nathan Srebro, TTI-C nati@tti-c.org > 834-7493 > > > _______________________________________________ > Colloquium mailing list > Colloquium@ttic.uchicago.edu > http://ttic.uchicago.edu/mailman/listinfo/colloquium > > From macglashan at tti-c.org Tue Apr 8 11:32:11 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Apr 8 12:32:09 2008 Subject: [TTIC Colloquium] Mark Braverman, University of Toronto- TTI-C Talk Message-ID: <005601c89996$1e1eac90$aabf8780@jmacglDPLFYD1> When: Friday, April 11 @ 10:00am Where: TTI-C Conference Room Who: Mark Braverman, University of Toronto Topic: Computability and Complexity of Julia sets Studying dynamical systems is key to understanding a wide range of phenomena ranging from planets' movement to climate patterns to market dynamics. Various numerical tools have been developed to address specific questions about dynamical systems, such as predicting the weather or planning the trajectory of a satellite. However, the theory of computation behind these problems appears to be very difficult to develop. While we have vast knowledge about computability and complexity of discrete problems, little is known about computability of even the most natural problems arising from dynamical systems. The focus of our study is dynamical systems that arise from iterating quadratic polynomials on the complex plane. They give rise to the amazing variety of fractals known as Julia sets, and are closely connected to the Mandelbrot set. Julia sets are perhaps the most drawn objects in Mathematics due to their fascinating fractal structure. The theory behind them is even more fascinating, and the dynamical systems generating them are in many ways archetypal. In this talk we discuss what it means for a planar set to be computable. We then present a variety of recent results, both positive and negative, on the computability and complexity of Julia sets. In particular we show that while the vast majority of Julia sets are computable -many even in polynomial time, some are as hard to compute as the Halting Problem and will never be drawn. The work paves the way to understanding computational properties of more complicated dynamical systems. Contact: Nikhil Devanur, TTI-C nikhil@tti-c.org 834-3541 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20080408/54d83cbf/attachment.htm From macglashan at tti-c.org Tue Apr 8 15:33:24 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Apr 8 16:33:16 2008 Subject: [TTIC Colloquium] K V Subrahmanyam (Chennai Mathematical Institute, INDIA)- TTI-C Talk Message-ID: <000301c899b7$cbbc7690$aabf8780@jmacglDPLFYD1> When: Thursday, April 10 @ 10:30am Where: TTI-C Conference Room Who: K. V. Subrahmanyam, Chennai Mathematical Institute, INDIA Topic: An introduction to Geometric Complexity Theory In this talk I will give an introduction to Geometric Complexity theory as an approach to separating computational complexity classes. I will talk about the algorithmic representation theory problems which arise in this approach. I will assume no background, and begin by introducing the important notions of completeness, in complexity theory, and class varieties, defined in GCT. I will illustrate the approach by considering the example of separating the permanent class from the determinant. Contact: Prahladh Harsha, TTI-C prahladh@tti-c.org 834-2549 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20080408/3c9ffab6/attachment-0001.htm From macglashan at tti-c.org Thu Apr 10 14:24:04 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Thu Apr 10 15:23:59 2008 Subject: [TTIC Colloquium] Mark Braverman, University of Toronto- TTI-C Talk Message-ID: <001b01c89b40$70ffe590$aabf8780@jmacglDPLFYD1> When: Friday, April 11 @ 10:00am Where: TTI-C Conference Room Who: Mark Braverman, University of Toronto Topic: Computability and Complexity of Julia sets Studying dynamical systems is key to understanding a wide range of phenomena ranging from planets' movement to climate patterns to market dynamics. Various numerical tools have been developed to address specific questions about dynamical systems, such as predicting the weather or planning the trajectory of a satellite. However, the theory of computation behind these problems appears to be very difficult to develop. While we have vast knowledge about computability and complexity of discrete problems, little is known about computability of even the most natural problems arising from dynamical systems. The focus of our study is dynamical systems that arise from iterating quadratic polynomials on the complex plane. They give rise to the amazing variety of fractals known as Julia sets, and are closely connected to the Mandelbrot set. Julia sets are perhaps the most drawn objects in Mathematics due to their fascinating fractal structure. The theory behind them is even more fascinating, and the dynamical systems generating them are in many ways archetypal. In this talk we discuss what it means for a planar set to be computable. We then present a variety of recent results, both positive and negative, on the computability and complexity of Julia sets. In particular we show that while the vast majority of Julia sets are computable -many even in polynomial time, some are as hard to compute as the Halting Problem and will never be drawn. The work paves the way to understanding computational properties of more complicated dynamical systems. Contact: Nikhil Devanur, TTI-C nikhil@tti-c.org 834-3541 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20080410/45926c0a/attachment.htm From macglashan at tti-c.org Mon Apr 14 09:33:43 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Apr 14 10:33:15 2008 Subject: [TTIC Colloquium] Ashok Veeraraghavan: University of Maryland, College Park- TTI-C Talk Message-ID: <000c01c89e3c$8a9758e0$aabf8780@jmacglDPLFYD1> When: Thursday, April 17 @ 10:30am Where: TTI-C Conference Room Who: Ashok Veeraraghavan, University of Maryland, College Park Topic: "Less is More" - Coded Computational Photography I will present the idea that a simple patterned attenuator (mask) acts as a very powerful modulator and allows us to come up with enhanced optical designs for alleviating many of the problems with traditional photography such as motion blur, focus blur and glare. I will show how a simple mask that can be printed on a transparency allows dramatic increase in the depth of field of the camera. Next, I will argue that alternative higher-dimensional representations of visual information (such as light-fields) allow for efficient and tractable algorithms for several traditionally hard vision problems such as depth estimation and separation of depth and texture edges. I will discuss how a simple high resolution mask allows the recovery of light-fields from a single captured image. I will also briefly touch upon another area of interest - Pattern recognition in video and show applications of such techniques to problems such as action recognition, simultaneous tracking and behavior analysis and video mining. The common link underlying these applications is the use of spatio-temporal information for robust estimation. 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/20080414/4d0ef27b/attachment.htm From macglashan at tti-c.org Wed Apr 16 09:12:59 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Wed Apr 16 10:12:30 2008 Subject: [TTIC Colloquium] Ashok Veeraraghavan: University of Maryland, College Park- TTI-C Talk Message-ID: <000e01c89fcb$fa835e70$aabf8780@jmacglDPLFYD1> When: Thursday, April 17 @ 10:30am Where: TTI-C Conference Room Who: Ashok Veeraraghavan, University of Maryland, College Park Topic: "Less is More" - Coded Computational Photography I will present the idea that a simple patterned attenuator (mask) acts as a very powerful modulator and allows us to come up with enhanced optical designs for alleviating many of the problems with traditional photography such as motion blur, focus blur and glare. I will show how a simple mask that can be printed on a transparency allows dramatic increase in the depth of field of the camera. Next, I will argue that alternative higher-dimensional representations of visual information (such as light-fields) allow for efficient and tractable algorithms for several traditionally hard vision problems such as depth estimation and separation of depth and texture edges. I will discuss how a simple high resolution mask allows the recovery of light-fields from a single captured image. I will also briefly touch upon another area of interest - Pattern recognition in video and show applications of such techniques to problems such as action recognition, simultaneous tracking and behavior analysis and video mining. The common link underlying these applications is the use of spatio-temporal information for robust estimation. 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/20080416/85f13119/attachment-0001.htm From macglashan at tti-c.org Fri Apr 18 14:12:45 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Fri Apr 18 15:12:02 2008 Subject: [TTIC Colloquium] Mihai Patrascu, MIT- TTI-C Talk Message-ID: <002c01c8a188$2f7eea50$aabf8780@jmacglDPLFYD1> When: Tuesday, April 22 @ 2:30pm Where: TTI-C Conference Room Who: Mihai Patrascu, MIT Topic: DATA STRUCTURES We show that a large fraction of the data-structure lower bounds known today in fact follow by reduction from the communication complexity of lopsided (asymmetric) set disjointness! This includes lower bounds for: * high-dimensional problems, where the goal is to show large space lower bounds. * constant-dimensional geometric problems, where the goal is to bound the query time for space O(n.polylg n). * dynamic problems, where we are looking for a trade-off between query and update time. (In this case, our bounds are slightly weaker than the originals, losing a lglgn factor.) Our reductions also imply new lower bounds for range reporting, the partial match problem, and reachability oracles. Contact: Prahladh Harsha, TTI-C prahladh@tti-c.org 834-2549 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20080418/6e8e3b84/attachment.htm From macglashan at tti-c.org Mon Apr 21 09:05:26 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Apr 21 10:04:42 2008 Subject: [TTIC Colloquium] Mihai Patrascu, MIT- TTI-C Talk Message-ID: <000701c8a3b8$c04582b0$aabf8780@jmacglDPLFYD1> When: Tuesday, April 22 @ 2:30pm Where: TTI-C Conference Room Who: Mihai Patrascu, MIT Topic: Data STRUCTURES We show that a large fraction of the data-structure lower bounds known today in fact follow by reduction from the communication complexity of lopsided (asymmetric) set disjointness! This includes lower bounds for: * high-dimensional problems, where the goal is to show large space lower bounds. * constant-dimensional geometric problems, where the goal is to bound the query time for space O(n.polylg n). * dynamic problems, where we are looking for a trade-off between query and update time. (In this case, our bounds are slightly weaker than the originals, losing a lglgn factor.) Our reductions also imply new lower bounds for range reporting, the partial match problem, and reachability oracles. Contact: Prahladh Harsha, TTI-C prahladh@tti-c.org 834-2549 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20080421/0c2a8a68/attachment.htm From macglashan at tti-c.org Tue Apr 22 09:17:54 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Apr 22 10:17:13 2008 Subject: [TTIC Colloquium] Nisheeth Vishnoi, IBM India Research Lab- TTI-C Talk In-Reply-To: <000701c8a3b8$c04582b0$aabf8780@jmacglDPLFYD1> References: <000701c8a3b8$c04582b0$aabf8780@jmacglDPLFYD1> Message-ID: <000501c8a483$adfe89f0$aabf8780@jmacglDPLFYD1> When: Tuesday, April 29 @ 10:00am Where: TTI-C Conference Room Who: Dr. Nisheeth K. Vishnoi, IBM India Research Lab Topic: Cuts, Embeddings, Flows and Unique Games A fundamental problem in computing is to cut a graph into two "roughly" equal parts so as to minimize the number of edges crossing the partition. Interest in this problem derives both from its numerous practical applications such as image segmentation, VLSI layout, parallel computing and clustering, and from its theoretical connections to a diverse set of powerful technical areas such as spectral methods, linear/semi-definite programming (SDP), measure concentration, metric embeddings, fourier analysis, probabilistically checkable proofs and the multiplicative weight update method. In this talk I will address theoretical and applied aspects concerning graph partitioning. >From an applied point of view, the main challenge is to design good, yet scalable algorithms for this problem. The reason being that, in practice, typical inputs to this problem consist of very large graphs, and hence, it is imperative to find algorithms that not only run very fast but provide a guarantee about the quality of the cut they produce. The first major step towards this was taken by Khandekar, Rao and Vazirani. They reduced the graph partitioning problem to the computation of a ``small'' number of max-flows and yet achieved an approximation factor of O(log ^2 n). In the first half of this talk, I will outline a combinatorial algorithm for graph partitioning which achieves an O(log n) approximation factor and runs in essentially max-flow time. >From a theoretical standpoint the main question is to determine approximability of this problem. For the upper bound, a result of Arora, Rao and Vazirani establishes an O(sqrt(logn)) approximation for this problem. For the lower bound, nothing much is known beyond NP-hardness. In light of almost no progress in proving hardness of approximation results, it would be a wishful conjecture to make that there is a constant factor approximation algorithm for graph partitioning. Indeed, an even stronger conjecture was made by the experts suggesting that, in fact, the ARV algorithm itself is a constant factor approximation algorithm. This conjecture also had, as ``close cousins'', certain well-studied and long standing conjectures in the theory of metric embeddings. In the second half of this talk, I will give a bird's eye view of the disproof of these conjectures and, time permitting, their connection to Unique Games. This talk is based on a series of joint works with Khot [FOCS 05], Devanur, Khot and Saket [STOC 06], Orecchia, Schulman, (Umesh) Vazirani [STOC 08], Arora, Khot, Kolla, Steurer, Tulsiani [STOC 08]. Contact: Julia Chuzhoy, TTI-C cjulia@tti-c.org 834-2490 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20080422/9a905b96/attachment-0001.htm From macglashan at tti-c.org Wed Apr 23 10:55:16 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Wed Apr 23 11:54:19 2008 Subject: [TTIC Colloquium] Sameer Shirdhonkar: University of Maryland, College Park- TTI-C Talk References: <000701c8a3b8$c04582b0$aabf8780@jmacglDPLFYD1> Message-ID: <004701c8a55a$6d14d2a0$aabf8780@jmacglDPLFYD1> When: Wednesday, April 30 @ 10:00am Where: TTI-C Conference Room Who: Sameer Shirdhonkar, University of Maryland, College Park Topic: Linear time approximation to the earth mover's distance for comparing histograms The earth mover's distance (EMD) or Vasershtein metric is a perceptually meaningful metric used for comparing histograms in various computer vision applications such as content based image retrieval. Computing the EMD involves solving a Kantorovich-Rubinstein (KR) transshipment problem and requires about cubic time. I will present a fast linear time approximation method for solving this class of KR transshipment problems. The approximation, called wavelet EMD (WEMD), is achieved by transforming the dual problem into the wavelet domain, where it admits an explicit solution. It is a weighted norm on the wavelet coefficients of the difference histograms and is equivalent to the EMD, i.e. the ratio of the two has lower and upper bounds. I will use colour histogram based image retrieval experiments to show that wavelet EMD is a good approximation to EMD with similar performance, but requires much less computation time. Briefly, I will talk about my previous work on the importance of non-negative lighting in model based recognition of specular objects. I will describe how to constrain the time domain values of a function when you can only manipulate its frequency domain coefficients. I will also demonstrate the Columbia-Maryland-Smithsonian electronic field guide for the identification of plant species using leaf images. This is joint work with my advisor Prof. David Jacobs. The electronic field guide is joint work with many people at the Columbia University, the University of Maryland and the Smithsonian Institution. 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/20080423/3043338c/attachment.htm From macglashan at tti-c.org Thu Apr 24 09:32:00 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Thu Apr 24 10:30:59 2008 Subject: [TTIC Colloquium] Tomaso Poggio (MIT) TTI-C Talk References: <000701c8a3b8$c04582b0$aabf8780@jmacglDPLFYD1> Message-ID: <001901c8a617$f53c7da0$aabf8780@jmacglDPLFYD1> When: Thursday, May 1 @ 10:30am Where: TTI-C Conference Room Who: Tomaso Poggio, Massachusetts Institute of Technology Topic: Models of Visual Recognition in the Ventral Stream: towards a theory We will describe a class of quantitative models of the ventral stream for object recognition, which have been developed during the last two decades from the anatomical and physiological data and which are quite successful in explaining several physiological data across different visual areas. Surprisingly, such models also mimic the level of human performance in difficult rapid image categorization tasks in which human vision is forced to operate in a feedforward mode. The main focus of the talk will be on extending these hierarchical models to deal with recognition in image sequences and to account for computational roles of attention and cortical backprojections. Finally we will also sketch initial steps with S. Smale, J. Bouvrie and L. Rosasco towards a mathematical theory of hierarchical architectures described by the models of visual cortex. Relevant papers can be downloaded from: http://cbcl.mit.edu/ Contact: Steve Smale, TTI-C smale@tti-c.org 834-2510 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20080424/60fcb599/attachment.htm From macglashan at tti-c.org Thu Apr 24 09:46:45 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Thu Apr 24 10:45:51 2008 Subject: [TTIC Colloquium] [Reformatted] Tomaso Poggio (MIT) TTI-C Talk Message-ID: <002c01c8a61a$06bb0720$aabf8780@jmacglDPLFYD1> When: ??Thursday, May 1 @ 10:30am Where:??TTI-C Conference Room Who:????Tomaso Poggio, Massachusetts Institute of Technology Topic:? Models of Visual Recognition in the Ventral Stream: towards a theory We will describe a class of quantitative models of the ventral stream for object recognition, which have been developed during the last two decades from the anatomical and physiological data and which are quite successful in explaining several physiological data across different visual areas. Surprisingly, such models also mimic the level of human performance in difficult rapid image categorization tasks in which human vision is forced to operate in a feedforward mode. The main focus of the talk will be on extending these hierarchical models to deal with recognition in image sequences and to account for computational roles of attention and cortical backprojections. Finally we will also sketch initial steps with S. Smale, J. Bouvrie and L. Rosasco towards a mathematical theory of hierarchical architectures described by the models of visual cortex. Relevant papers can be downloaded from:? http://cbcl.mit.edu/ Contact:?????Steve Smale, TTI-C?????smale@tti-c.org?????834-2510 From macglashan at tti-c.org Mon Apr 28 09:21:47 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Apr 28 10:20:30 2008 Subject: [TTIC Colloquium] Nisheeth Vishnoi, IBM India Research Lab- TTI-C Talk References: <000701c8a3b8$c04582b0$aabf8780@jmacglDPLFYD1> Message-ID: <000901c8a93b$31ee6210$aabf8780@jmacglDPLFYD1> When: Tuesday, April 29 @ 10:00am Where: TTI-C Conference Room Who: Dr. Nisheeth K. Vishnoi, IBM India Research Lab Topic: Cuts, Embeddings, Flows and Unique Games A fundamental problem in computing is to cut a graph into two "roughly" equal parts so as to minimize the number of edges crossing the partition. Interest in this problem derives both from its numerous practical applications such as image segmentation, VLSI layout, parallel computing and clustering, and from its theoretical connections to a diverse set of powerful technical areas such as spectral methods, linear/semi-definite programming (SDP), measure concentration, metric embeddings, fourier analysis, probabilistically checkable proofs and the multiplicative weight update method. In this talk I will address theoretical and applied aspects concerning graph partitioning. >From an applied point of view, the main challenge is to design good, yet scalable algorithms for this problem. The reason being that, in practice, typical inputs to this problem consist of very large graphs, and hence, it is imperative to find algorithms that not only run very fast but provide a guarantee about the quality of the cut they produce. The first major step towards this was taken by Khandekar, Rao and Vazirani. They reduced the graph partitioning problem to the computation of a ``small'' number of max-flows and yet achieved an approximation factor of O(log ^2 n). In the first half of this talk, I will outline a combinatorial algorithm for graph partitioning which achieves an O(log n) approximation factor and runs in essentially max-flow time. >From a theoretical standpoint the main question is to determine approximability of this problem. For the upper bound, a result of Arora, Rao and Vazirani establishes an O(sqrt(logn)) approximation for this problem. For the lower bound, nothing much is known beyond NP-hardness. In light of almost no progress in proving hardness of approximation results, it would be a wishful conjecture to make that there is a constant factor approximation algorithm for graph partitioning. Indeed, an even stronger conjecture was made by the experts suggesting that, in fact, the ARV algorithm itself is a constant factor approximation algorithm. This conjecture also had, as ``close cousins'', certain well-studied and long standing conjectures in the theory of metric embeddings. In the second half of this talk, I will give a bird's eye view of the disproof of these conjectures and, time permitting, their connection to Unique Games. This talk is based on a series of joint works with Khot [FOCS 05], Devanur, Khot and Saket [STOC 06], Orecchia, Schulman, (Umesh) Vazirani [STOC 08], Arora, Khot, Kolla, Steurer, Tulsiani [STOC 08]. Contact: Julia Chuzhoy, TTI-C cjulia@tti-c.org 834-2490 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20080428/7b4fccfe/attachment-0001.htm From macglashan at tti-c.org Tue Apr 29 08:59:06 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Apr 29 09:57:59 2008 Subject: [TTIC Colloquium] Sameer Shirdhonkar: University of Maryland, College Park- TTI-C Talk References: <000701c8a3b8$c04582b0$aabf8780@jmacglDPLFYD1> Message-ID: <000c01c8aa01$36ad10a0$aabf8780@jmacglDPLFYD1> When: Wednesday, April 30 @ 10:00am Where: TTI-C Conference Room, 1427 E. 60th St. Who: Sameer Shirdhonkar, University of Maryland, College Park Topic: Linear time approximation to the earth mover's distance for comparing histograms The earth mover's distance (EMD) or Vasershtein metric is a perceptually meaningful metric used for comparing histograms in various computer vision applications such as content based image retrieval. Computing the EMD involves solving a Kantorovich-Rubinstein (KR) transshipment problem and requires about cubic time. I will present a fast linear time approximation method for solving this class of KR transshipment problems. The approximation, called wavelet EMD (WEMD), is achieved by transforming the dual problem into the wavelet domain, where it admits an explicit solution. It is a weighted norm on the wavelet coefficients of the difference histograms and is equivalent to the EMD, i.e. the ratio of the two has lower and upper bounds. I will use colour histogram based image retrieval experiments to show that wavelet EMD is a good approximation to EMD with similar performance, but requires much less computation time. Briefly, I will talk about my previous work on the importance of non-negative lighting in model based recognition of specular objects. I will describe how to constrain the time domain values of a function when you can only manipulate its frequency domain coefficients. I will also demonstrate the Columbia-Maryland-Smithsonian electronic field guide for the identification of plant species using leaf images. This is joint work with my advisor Prof. David Jacobs. The electronic field guide is joint work with many people at the Columbia University, the University of Maryland and the Smithsonian Institution. 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/20080429/18041b09/attachment.htm From macglashan at tti-c.org Wed Apr 30 10:59:02 2008 From: macglashan at tti-c.org (Julia MacGlashan) Date: Wed Apr 30 11:57:41 2008 Subject: [TTIC Colloquium] Tomaso Poggio (MIT) TTI-C Talk Message-ID: <002a01c8aadb$1cc10270$aabf8780@jmacglDPLFYD1> When: Thursday, May 1 @ 10:30am Where: TTI-C Conference Room Who: Tomaso Poggio, Massachusetts Institute of Technology Topic: Models of Visual Recognition in the Ventral Stream: towards a theory We will describe a class of quantitative models of the ventral stream for object recognition, which have been developed during the last two decades from the anatomical and physiological data and which are quite successful in explaining several physiological data across different visual areas. Surprisingly, such models also mimic the level of human performance in difficult rapid image categorization tasks in which human vision is forced to operate in a feedforward mode. The main focus of the talk will be on extending these hierarchical models to deal with recognition in image sequences and to account for computational roles of attention and cortical backprojections. Finally we will also sketch initial steps with S. Smale, J. Bouvrie and L. Rosasco towards a mathematical theory of hierarchical architectures described by the models of visual cortex. Relevant papers can be downloaded from: http://cbcl.mit.edu/ Contact: Steve Smale, TTI-C smale@tti-c.org 834-2510 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20080430/3c986240/attachment.htm