From cnovak at tti-c.org Mon Mar 2 08:51:37 2009 From: cnovak at tti-c.org (Christina Novak) Date: Mon Mar 2 09:14:09 2009 Subject: [TTIC Colloquium] TTI-C Distinguished Lecture Series (March 3, Dr. Mihalis Yannakakis- Columbia University) Message-ID: Good morning, Please join us tomorrow for TTI-C's second 2009 Distinguished Lecture Series speaker, Dr. Mihalis Yannakakis. The talk is from 2 pm, and located in TTI-C's new facility at 6045 S. Kenwood Ave (and 61st St.). More details, speaker bio and directions may be found at http://tti-c.org/dls. We look forward to seeing you at the lecture! Sincerely, David McAllester TTI-C Chief Academic Officer For questions about the TTI-C Distinguished Lecture Series, please contact: Chrissy Novak cnovak@tti-c.org or (773)834-2216. Tuesday, March 3rd, 2009 (2 pm) Mihalis Yannakakis (Columbia University) "Equilibria, Fixed Points, and Complexity Classes" Abstract: Many models from a variety of areas involve the computation of an equilibrium or fixed point of some kind. Examples include Nash equilibria in games; market equilibria; computing optimal strategies and the values of competitive games (stochastic and other games); stable configurations of neural networks; analyzing basic stochastic models for evolution like branching processes and for language like stochastic context-free grammars; and models that incorporate the basic primitives of probability and recursion like recursive Markov chains. It is not known whether these problems can be solved in polynomial time. Despite their broad diversity, there are certain common computational principles that underlie different types of equilibria and connect many of these problems to each other. In this talk we will discuss these common principles and the corresponding complexity classes that capture them. -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090302/873fd220/attachment.htm From macglashan at tti-c.org Mon Mar 2 09:42:11 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Mar 2 09:27:09 2009 Subject: [TTIC Colloquium] TTI-C Colloquium: Ohad Shamir, Hebrew University of Jerusalem References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <424313679C2D401399620F4A3E370542@jmacglDPLFYD1> When: Monday, March 9 @ 2:00pm Where: TTI-C Conference Room #526, 6045 S Kenwood Ave, 5th Floor Who: Ohad Shamir (Hebrew University of Jerusalem) Title: Vox Populi: On Learning from Crowds With the emergence of large scale search engines and crowd-sourcing websites, machine learning practitioners increasingly have to cope with datasets labeled by a large number of teachers. Unlike traditional datasets, which are usually labeled by a small select group of experts, here no a-priori knowledge is available on the teachers, and their quality usually ranges from excellent to downright malicious. Often, there is no control over the assignment of instances to teachers, and common multi-teacher techniques can be impossible or very wasteful to imple-ment. As a result, these kind of datasets pose new challenges to both the theory and the practice of machine learning. In this talk, I'll discuss some novel approaches to improve and learn from such data, which use the collective "wisdom of the crowd" to mitigate the effect of low quality and malicious teachers. Joint work with Ofer Dekel Contact: Shai Shalev-Shwartz, TTI-C shai@tti-c.org 834-6850 -------------- next part -------------- A non-text attachment was scrubbed... Name: winmail.dat Type: application/ms-tnef Size: 7978 bytes Desc: not available Url : http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090302/af722f59/winmail.bin From macglashan at tti-c.org Tue Mar 3 08:12:02 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Mar 3 07:57:00 2009 Subject: [TTIC Colloquium] TTI-C Talk: Tamir Hazan (Hebrew University of Jerusalem) References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <023A83FFF192445C8A3E65E56BEB137D@jmacglDPLFYD1> REMINDER When: Wednesday, March 4 @ 11:00am (lunch will be provided after talk) Where: TTI-C Conference Room #526, 6045 S Kenwood Ave, 5th Floor Who: Tamir Hazan (Hebrew University of Jerusalem) Title: Convex Belief Propagation - Approximated inference and LP-relaxations We derive a one-parameter local message-passing algorithm, called "norm-product", which covers both the tasks of computing approximate marginal probabilities and maximum a posteriori (MAP) assignment for general graphical models. A parameter $\epsilon$ controls a perturbation term of a "fractional entropy approximation" $\tilde H$ which includes Bethe, Tree-reweighted (TRW) and convex entropy approximations. When $\tilde H$ is the Bethe approximation, the settings $\epsilon=0$ and $\epsilon=1$ produce the max-product and sum-product algorithms, respectively. When $\tilde H$ is a convex entropy approximation and $\epsilon\rightarrow 0$, the algorithm is a globally convergent Linear Programming (LP) relaxation of the MAP problem. When $\tilde H$ is convex and $\epsilon=1$, norm-product is a globally convergent algorithm for "convex free energies" for approximate marginal probabilities, and when $\epsilon=0$ norm-product becomes a family of convergent "max-product-like" algorithms for computing approximate MAP. Contact: Shai Shalev-Shwartz, TTI-C shai@tti-c.org 834-6850 -------------- next part -------------- A non-text attachment was scrubbed... Name: winmail.dat Type: application/ms-tnef Size: 7978 bytes Desc: not available Url : http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090303/6e570e43/winmail-0001.bin From macglashan at tti-c.org Wed Mar 4 09:27:49 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Wed Mar 4 09:12:43 2009 Subject: [TTIC Colloquium] TTI-C Talk: Satyen Kale (Microsoft) References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <091C9070DA044112A9B416B28FE42F5E@jmacglDPLFYD1> REMINDER When: Thursday, March 5 @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Satyen Kale (Microsoft Research New England) Title: Improved Decision-Making under Uncertainty Decision-making in the face of uncertainty over future outcomes is a fundamental algorithmic task, with roots in statistics and information theory, and applications in machine learning, signal processing, network routing and finance. The framework of regret minimization captures the notion of online decision-making algorithms that are competitive with the best possible decision in hindsight, under minimal assumptions on how the costs of the decisions are set. A major achievement of online learning theory has been the development of algorithms that minimize regret even under such "worst-case" assumptions. However, the regret bounds are quite suboptimal in real-life scenarios where the decision costs are particularly benign. On the other hand, "average-case" learning methods that posit a specific stochastic model on the costs have better convergence bounds, but may fail to work when the actual costs deviate from the model. Although these algorithms have been developed over several decades, designing regret minimizing algorithms that smoothly handle benign cost sequences, while not compromising worst-case robustness, was considered a significant open problem. In my talk, I will give an overview of my recent work with Elad Hazan which solved this open problem for four fundamental online learning scenarios: (a) prediction from expert advice, (b) online linear optimization, (c) universal portfolio selection, and (d) bandit linear optimization. The work on portfolio selection algorithms has implications in the standard Brownian motion model of stock prices, which were verified by experiments on real data. Contact: Sham Kakade, TTI-C sham@tti-c.org 834-2550 -------------- next part -------------- A non-text attachment was scrubbed... Name: winmail.dat Type: application/ms-tnef Size: 8394 bytes Desc: not available Url : http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090304/ee7e5f09/winmail.bin From macglashan at tti-c.org Thu Mar 5 09:38:01 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Thu Mar 5 09:22:49 2009 Subject: [TTIC Colloquium] TTI-C Talk: Sanghyun Park (Argonne National Laboratory) References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <4D900CD05C1F418FA5844888B4813E2D@jmacglDPLFYD1> When: Tuesday, March 10th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Sanghyun Park (Argonne National Laboratory) Title: Physics-Based Simulations of Protein Conformations Many important phenomena in molecular biology, such as protein folding, protein-ligand binding, and protein allostery, involve conformational changes of proteins. Physics-based simulations are commonly used for the study of protein conformations, but many problems are still out of reach of such simulations and call for original algorithmic solutions. In this talk, I will lay out some of the theoretical challenges and present my new method "deactivated morphing" that is robustly applicable to conformational changes of arbitrary complexity. Contact: Jinbo Xu, TTI-C j3xu@tti-c.org 834-2511 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090305/9770fcbf/attachment.htm From macglashan at tti-c.org Thu Mar 5 11:32:42 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Thu Mar 5 11:17:30 2009 Subject: [TTIC Colloquium] TTI-C Talk: Daniel Spoonhower, CMU References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <82B4779958A24061A53207B6F446FA51@jmacglDPLFYD1> When: Wednesday, March 11th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Daniel Spoonhower (Carnegie Mellon University) Title: Scheduling Deterministic Parallel Programs Deterministic parallelism enables programmers to write programs without concern for how parallel tasks are interleaved. While this simplifies reasoning about the correctness of parallel programs, the performance of these programs still depends on many aspects of the language implementation, including the scheduling policy. For example, the choice of scheduling policy can asymptotically increase the amount of memory required to run an application. In this talk, I will give some background on scheduling and present a methodology for understanding the performance of parallel programs. At the core of this work is a cost semantics that enables programmers to reason formally about different scheduling policies and how they affect performance. This cost semantics is the basis for a suite of prototype profiling tools. These tools enable programmers to simulate and visualize program execution under different scheduling policies. My cost semantics also provides a specification for an implementation of the language. I have extended MLton, a compiler for Standard ML, with support for parallelism and implemented several different scheduling policies. Using my cost semantics and profiler, I found a memory leak caused by a bug in one of the existing optimizations in MLton. I will also talk about how this implementation has inspired some theoretical work on parallel scheduling. Though work stealing has been successfully applied to nested parallelism, I will present new bounds on the overhead of work stealing implementations of parallel futures, a more expressive form of parallelism. These bounds apply to all prior work stealing schedulers but suggest the possibility of more efficient alternatives. Contact: Umut Acar, TTI-C umut@tti-c.org 702-5072 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090305/54e78673/attachment-0001.htm From macglashan at tti-c.org Fri Mar 6 09:37:34 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Fri Mar 6 09:22:22 2009 Subject: [TTIC Colloquium] TTI-C Colloquium: Ohad Shamir, Hebrew University of Jerusalem References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <7EEC77E1A6CA45D99648CF3E23164988@jmacglDPLFYD1> REMINDER When: Monday, March 9 @ 2:00pm Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Ohad Shamir (Hebrew University of Jerusalem) Title: Vox Populi: On Learning from Crowds With the emergence of large scale search engines and crowd-sourcing websites, machine learning practitioners increasingly have to cope with datasets labeled by a large number of teachers. Unlike traditional datasets, which are usually labeled by a small select group of experts, here no a-priori knowledge is available on the teachers, and their quality usually ranges from excellent to downright malicious. Often, there is no control over the assignment of instances to teachers, and common multi-teacher techniques can be impossible or very wasteful to imple-ment. As a result, these kind of datasets pose new challenges to both the theory and the practice of machine learning. In this talk, I'll discuss some novel approaches to improve and learn from such data, which use the collective "wisdom of the crowd" to mitigate the effect of low quality and malicious teachers. Joint work with Ofer Dekel. Contact: Shai Shalev-Shwartz, TTI-C shai@tti-c.org 834-6850 -------------- next part -------------- A non-text attachment was scrubbed... Name: winmail.dat Type: application/ms-tnef Size: 8078 bytes Desc: not available Url : http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090306/80c05eb5/winmail.bin From macglashan at tti-c.org Mon Mar 9 08:39:50 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Mar 9 09:39:31 2009 Subject: [TTIC Colloquium] TTI-C Talk: Sanghyun Park (Argonne National Laboratory) References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <4A76D94C16E2451FAFCE99BBE27099D5@jmacglDPLFYD1> REMINDER When: Tuesday, March 10th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Sanghyun Park (Argonne National Laboratory) Title: Physics-Based Simulations of Protein Conformations Many important phenomena in molecular biology, such as protein folding, protein-ligand binding, and protein allostery, involve conformational changes of proteins. Physics-based simulations are commonly used for the study of protein conformations, but many problems are still out of reach of such simulations and call for original algorithmic solutions. In this talk, I will lay out some of the theoretical challenges and present my new method "deactivated morphing" that is robustly applicable to conformational changes of arbitrary complexity. Contact: Jinbo Xu, TTI-C j3xu@tti-c.org 834-2511 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090309/150ce200/attachment.htm From macglashan at tti-c.org Tue Mar 10 08:25:38 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Mar 10 09:25:53 2009 Subject: [TTIC Colloquium] Vision Seminar: Gilad Lerman, University of Minnesota References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <175EF13A76074317AF6A3AE43B7B0B37@jmacglDPLFYD1> When: Wednesday, March 18th @ 1:00pm Where: 6045 S Kenwood Ave, TTI-C Conference Room #530 (5th Floor) Who: Gilad Lerman (University of Minnesota, Dept. of Mathematics) Title: Multi-manifold data modeling via spectral curvature clustering We propose a fast multi-way spectral clustering algorithm for multi-manifold data modeling, i.e., modeling data by mixtures of manifolds (possibly intersecting). We describe the supporting theory as well as the practical choices guided by it. We first develop the case of hybrid linear modeling, i.e., when the underlying manifolds are affine subspaces in a Euclidean space, and then we extend this setting to more general manifolds. We exemplify the practical use of the algorithm by demonstrating its successful application to problems of motion segmentation. Contact: Ronen Basri, TTI-C ronen.basri@tti-c.org 834-2515 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090310/813b70a8/attachment-0001.htm From macglashan at tti-c.org Tue Mar 10 08:28:57 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Mar 10 09:29:11 2009 Subject: [TTIC Colloquium] TTI-C Talk: Daniel Spoonhower, CMU References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <3AC5A6855CD2469D8D855D6815DE112D@jmacglDPLFYD1> REMINDER When: Wednesday, March 11th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Daniel Spoonhower (Carnegie Mellon University) Title: Scheduling Deterministic Parallel Programs Deterministic parallelism enables programmers to write programs without concern for how parallel tasks are interleaved. While this simplifies reasoning about the correctness of parallel programs, the performance of these programs still depends on many aspects of the language implementation, including the scheduling policy. For example, the choice of scheduling policy can asymptotically increase the amount of memory required to run an application. In this talk, I will give some background on scheduling and present a methodology for understanding the performance of parallel programs. At the core of this work is a cost semantics that enables programmers to reason formally about different scheduling policies and how they affect performance. This cost semantics is the basis for a suite of prototype profiling tools. These tools enable programmers to simulate and visualize program execution under different scheduling policies. My cost semantics also provides a specification for an implementation of the language. I have extended MLton, a compiler for Standard ML, with support for parallelism and implemented several different scheduling policies. Using my cost semantics and profiler, I found a memory leak caused by a bug in one of the existing optimizations in MLton. I will also talk about how this implementation has inspired some theoretical work on parallel scheduling. Though work stealing has been successfully applied to nested parallelism, I will present new bounds on the overhead of work stealing implementations of parallel futures, a more expressive form of parallelism. These bounds apply to all prior work stealing schedulers but suggest the possibility of more efficient alternatives. Contact: Umut Acar, TTI-C umut@tti-c.org 702-5072 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090310/36257685/attachment.htm From macglashan at tti-c.org Wed Mar 11 09:24:07 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Wed Mar 11 10:25:00 2009 Subject: [TTIC Colloquium] TTI-C Talk: Jennifer Wortman, UPenn References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <80A7A1E3E4EA4891AA70487386EABC9D@jmacglDPLFYD1> When: Tuesday, March 17th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Jennifer Wortman, University of Pennsylvania Title: Learning from Collective Preferences, Behavior, and Beliefs Machine learning has become one of the most active and exciting areas of computer science research, in large part because of its wide-spread applicability to problems as diverse as natural language processing, speech recognition, spam detection, search, computer vision, gene discovery, medical diagnosis, and robotics. At the same time, the growing popularity of the Internet and social networking sites like Facebook has led to the availability of novel sources of data on the preferences, behavior, and beliefs of massive populations of users. Naturally, both researchers and engineers are eager to apply techniques from machine learning in order to aggregate and make sense of this wealth of collective information. However, traditional theories of learning fail to capture the complex issues that arise in such settings, and as a result, many of the techniques currently employed are ad hoc and not well understood. A major goal of my research is to narrow this gap between theory and practice by designing new learning models and algorithms to address and illuminate problems commonly faced when aggregating local information from large populations of users. In this talk, I will discuss two specific pieces of work that fall into this category. In the first, we develop a forecaster that is guaranteed to perform reasonably well compared to the best expert in a population but simultaneously never any worse than the average. In the second, we investigate the computational complexity of pricing in prediction markets, betting markets designed to aggregate individuals' beliefs about the likelihood of future events, and propose an approximation technique based on the previously unexplored connection between prediction market prices and learning from expert advice. Contact: Sham Kakade, TTI-C sham@tti-c.org 834-2550 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090311/377561c0/attachment.htm From macglashan at tti-c.org Thu Mar 12 09:20:49 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Thu Mar 12 10:22:22 2009 Subject: [TTIC Colloquium] New Course: Large Scale Learning, CMSC 35900 Message-ID: NEW COURSE: CMSC 35900 - Large Scale Learning, Spring 2009 Instructors: Sham Kakade (sham@tti-c.org), Greg Shakhnarovich (greg@tti-c.org) Time and place: TT 1:30-2:50pm, TTI-C room 530, 6045 S. Kenwood Ave., 5th floor Course website: http://ttic.uchicago.edu/~gregory/courses/LargeScaleLearning/ The course will focus on theory and practice of working with large data sets, characterized by large numbers of data points and/or high dimensions. We will cover a set of computational tools that allow efficient storage, search and inference in such data sets, as well as theoretical results pertaining to these tools. We consider both statistical and computational efficiency and limitations of the methods covered, and discuss data structures and algorithms for implementing these methods in practice. The major topics discussed in the course will include the following: * Dimensionality reduction, including random projections, spectral methods and embeddings. * Efficient indexing and search, including locality-sensitive hashing, metric trees * Large-scale non-parametric methods, including locally-weighted regression, example-based density estimation * Unsupervised statistical learning tasks, including clustering -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090312/9e096caa/attachment.htm From macglashan at tti-c.org Thu Mar 12 10:33:57 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Thu Mar 12 11:35:27 2009 Subject: [TTIC Colloquium] TTI-C Talk: Slav Petrov, UC Berkeley References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <9553A37119AC4B00AC502FDD0DDFD26C@jmacglDPLFYD1> When: Wednesday, March 18th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Slav Petrov (University of California, Berkeley) Title: Coarse-to-Fine Methods for Natural Language Processing State-of-the-art NLP models are anything but compact. Parsers have huge grammars, machine translation systems have huge transfer tables, and so on across a range of tasks. With such complexity come two challenges. First, how can we learn highly complex models? Second, how can we efficiently infer optimal structures within them? Hierarchical coarse-to-fine methods address both questions. Coarse-to-fine approaches exploit a sequence of models which introduce complexity gradually. At the top of the sequence is a trivial model in which learning and inference are both cheap. Each subsequent model refines the previous one, until a final, full-complexity model is reached. Because each refinement introduces only limited complexity, both learning and inference can be done in an incremental fashion. In this talk, I describe several coarse-to-fine systems. In the domain of syntactic parsing, complexity is in the grammar. I present a latent-variable approach which begins with an X-bar grammar and learns to iteratively refine grammar categories. For example, noun phrases might be split into subcategories for subjects and objects, singular and plural, and so on. This splitting process admits an efficient incremental inference scheme which reduces parsing times by orders of magnitude. This approach produces the best parsing accuracies across an array of languages, in a fully language-general fashion. In the domain of syntactic machine translation, complexity arises because there and too many target language word types. To manage this complexity, we translate into target language clusterings of increasing vocabulary size. This approach gives dramatic speed-ups while actually increasing final translation quality. Contact: Karen Livescu, TTI-C klivescu@tti-c.org 834-2549 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090312/7a700ac6/attachment-0001.htm From macglashan at tti-c.org Thu Mar 12 12:46:54 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Thu Mar 12 13:48:37 2009 Subject: [TTIC Colloquium] UC Math 2009 Unni Namboodiri Lectures Message-ID: <955091BFBDBB4D5FADDB56CDB81D108F@jmacglDPLFYD1> University of Chicago, Department of Mathematics announces the 2009 UNNI NAMBOODIRI LECTURES by MLADEN BESTVINA, University of Utah THREE LECTURES ON: The Topology and Geometry of OUT (Fn) Lecture I: SL2(Z) and Three Generalizations Tuesday, March 31, 4:30 - Room 206 Eckhart Hall - 1118 E. 58th Street Lecture II: Outer Space and the Topology of Out(Fn) Wednesday, April 1, 4:00 - Room 206 Eckhart Hall - 1118 E. 58th Street Lecture III: What is the Geometry of Outer Space Thursday, April 2, 4:30 - Room 206 Eckhart Hall - 1118 E. 58th Street Tea will be served in the Common Room, Eckhart 209, 30 minutes before each lecture. ___________________________________________________________________________ Unni Namboodiri (1956-1981), a brilliant student of mathematics at the University of Chicago under Professor J. Peter May, died in an automobile accident in December 1981, just a few days after presenting and defending his doctoral thesis, Equivariant vector fields on spheres. The Unni Namboodiri Lecture Series has been established in his memory by his family. Persons with a disability who believe they may need assistance please call Lynette Whalum at (773) 702-7100. -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090312/d658d89f/attachment.htm From macglashan at tti-c.org Fri Mar 13 12:32:18 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Fri Mar 13 13:34:36 2009 Subject: [TTIC Colloquium] TTI-C Talk: Leonid Sigal, University of Toronto References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <2BB750E78D5E43FD9F7852E08F2A5D7E@jmacglDPLFYD1> When: Thursday, March 19th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Leonid Sigal, University of Toronto Title: Physics-based Models and Priors for Human Motion Tracking Recovery and analysis of human pose and motion from video is the key enabling technology for a broad spectrum of applications, in and outside of computer science; including applications in HCI, biometrics, biomechanics and computer graphics. Despite years of research, the general problem of tracking a person in an unconstrained environment, particularly from monocular observations, remains challenging. In this talk I will describe the basic building blocks and challenges of the articulated human pose estimation and tracking, as well as my contributions to the various aspects of this problem and the field in recent years. I will particularly focus on the new and unique class of models that incorporate physic-based predictions and simulation into the inference process. Physics plays an important and intricate role in characterizing, describing and predicting human motion. The key benefit of using physics-based models or priors for tracking is the improved realism in the recovered motions, as well as enhanced ability to deal with weak image observations and diverse environmental interactions. Newtonian physics, in these models, approximates the rigid-body dynamics of the body in the environment through the application and integration of forces. Since the motion of the body is intimately tied with the environment, the use of such models also allows one to start reasoning about the geometry and physical properties of the environment as a whole (e.g. orientation and compliance of ground). This work is part of joint projects with colleagues at Brown University and University of Toronto. Bio: Leonid Sigal is a postdoctoral fellow in the Department of Computer Science at University of Toronto. He received his Ph.D. in computer science from Brown University (2007); his M.S. from Brown University (2003); his M.A. from Boston University (1999); and his B.Sc. degrees in Computer Science and Mathematics from Boston University (1999). From 1999 to 2001, he worked as a senior vision engineer at Cognex Corporation, where he developed industrial vision applications for pattern analysis and verification. In 2002, he spent a semester as a research intern at Siemens Corporate Research (SCR) working on autonomous obstacle detection and avoidance for vehicle navigation. During the summers of 2005 and 2006, he worked as a research intern at Intel Applications Research Lab (ARL) on human pose estimation and tracking. His work received the Best Paper Award at the Articulate Motion and Deformable Objects Conference in 2006 (with Prof. Michael J. Black). Dr. Sigal's research interests mainly lie in the areas of computer vision and machine learning, but also borderline fields of computer graphics, psychology and humanoid robotics. He is particularly interested in statistical models for problems of visual inference, including human motion recovery and analysis, graphical models, probabilistic and hierarchical inference. 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/20090313/6b6f2523/attachment.htm From macglashan at tti-c.org Mon Mar 16 10:56:49 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Mar 16 12:01:03 2009 Subject: [TTIC Colloquium] TTI-C Talk: Jennifer Wortman, UPenn References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <045C187F9E5A49699962E545F9DF20CF@jmacglDPLFYD1> REMINDER When: Tuesday, March 17th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Jennifer Wortman, University of Pennsylvania Title: Learning from Collective Preferences, Behavior, and Beliefs Machine learning has become one of the most active and exciting areas of computer science research, in large part because of its wide-spread applicability to problems as diverse as natural language processing, speech recognition, spam detection, search, computer vision, gene discovery, medical diagnosis, and robotics. At the same time, the growing popularity of the Internet and social networking sites like Facebook has led to the availability of novel sources of data on the preferences, behavior, and beliefs of massive populations of users. Naturally, both researchers and engineers are eager to apply techniques from machine learning in order to aggregate and make sense of this wealth of collective information. However, traditional theories of learning fail to capture the complex issues that arise in such settings, and as a result, many of the techniques currently employed are ad hoc and not well understood. A major goal of my research is to narrow this gap between theory and practice by designing new learning models and algorithms to address and illuminate problems commonly faced when aggregating local information from large populations of users. In this talk, I will discuss two specific pieces of work that fall into this category. In the first, we develop a forecaster that is guaranteed to perform reasonably well compared to the best expert in a population but simultaneously never any worse than the average. In the second, we investigate the computational complexity of pricing in prediction markets, betting markets designed to aggregate individuals' beliefs about the likelihood of future events, and propose an approximation technique based on the previously unexplored connection between prediction market prices and learning from expert advice. Contact: Sham Kakade, TTI-C sham@tti-c.org 834-2550 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090316/1f77aba8/attachment-0001.htm From macglashan at tti-c.org Mon Mar 16 14:43:38 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Mar 16 15:47:53 2009 Subject: [TTIC Colloquium] TTI-C Talk: Joseph Keshet, IDIAP Research Institute References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: When: Monday, March 23rd @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Joseph Keshet (IDIAP Research Institute, Switzerland) Title: Discriminative Spoken keyword Detection The current state-of-the-art automatic speech recognizers are mostly based on hidden Markov models (HMMs). Despite their popularity, HMM-based approaches have several known drawbacks such as training objective which is not aimed at optimizing the evaluation objective. We proposes a new approach for spoken keyword spotting, which is based on large margin and kernel methods rather than on HMMs. Unlike previous approaches, the proposed method employs a discriminative learning procedure, in which the learning phase aims at achieving a high area under the ROC curve, as this quantity is the most common measure to evaluate keyword spotters. The keyword spotter we devise is based on mapping the input acoustic representation of the speech utterance along with the target keyword into a vector space. Building on techniques used for large margin and kernel methods for predicting whole sequences, our keyword spotter distills to a classifier in this vector-space, which separates speech utterances in which the keyword is uttered from speech utterances in which the keyword is not uttered. We describe a simple iterative algorithm for training the keyword spotter and discuss its formal properties, showing theoretically that it attains high area under the ROC curve. Experimental results suggest that on variety standard speech recognition datasets our discriminative system outperforms the conventional context-independent HMM-based system. Contact: Karen Livescu, TTI-C klivescu@tti-c.org 834-2549 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090316/38ea1fce/attachment.htm From macglashan at tti-c.org Tue Mar 17 07:50:35 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Mar 17 08:55:11 2009 Subject: [TTIC Colloquium] Vision Seminar: Gilad Lerman, University of Minnesota References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <4E828A779DEF4A61AD8025A65F2642BB@jmacglDPLFYD1> REMINDER When: Wednesday, March 18th @ 1:00pm Where: 6045 S Kenwood Ave, TTI-C Conference Room #530 (5th Floor) Who: Gilad Lerman (University of Minnesota, Dept. of Mathematics) Title: Multi-manifold data modeling via spectral curvature clustering We propose a fast multi-way spectral clustering algorithm for multi-manifold data modeling, i.e., modeling data by mixtures of manifolds (possibly intersecting). We describe the supporting theory as well as the practical choices guided by it. We first develop the case of hybrid linear modeling, i.e., when the underlying manifolds are affine subspaces in a Euclidean space, and then we extend this setting to more general manifolds. We exemplify the practical use of the algorithm by demonstrating its successful application to problems of motion segmentation. Contact: Ronen Basri, TTI-C ronen.basri@tti-c.org 834-2515 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090317/ea011cde/attachment.htm From macglashan at tti-c.org Tue Mar 17 08:30:28 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Mar 17 09:34:58 2009 Subject: [TTIC Colloquium] TTI-C Talk: Slav Petrov, UC Berkeley References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <8E5BC6C397AD44BDB1AAE500A39B4C34@jmacglDPLFYD1> REMINDER When: Wednesday, March 18th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Slav Petrov (University of California, Berkeley) Title: Coarse-to-Fine Methods for Natural Language Processing State-of-the-art NLP models are anything but compact. Parsers have huge grammars, machine translation systems have huge transfer tables, and so on across a range of tasks. With such complexity come two challenges. First, how can we learn highly complex models? Second, how can we efficiently infer optimal structures within them? Hierarchical coarse-to-fine methods address both questions. Coarse-to-fine approaches exploit a sequence of models which introduce complexity gradually. At the top of the sequence is a trivial model in which learning and inference are both cheap. Each subsequent model refines the previous one, until a final, full-complexity model is reached. Because each refinement introduces only limited complexity, both learning and inference can be done in an incremental fashion. In this talk, I describe several coarse-to-fine systems. In the domain of syntactic parsing, complexity is in the grammar. I present a latent-variable approach which begins with an X-bar grammar and learns to iteratively refine grammar categories. For example, noun phrases might be split into subcategories for subjects and objects, singular and plural, and so on. This splitting process admits an efficient incremental inference scheme which reduces parsing times by orders of magnitude. This approach produces the best parsing accuracies across an array of languages, in a fully language-general fashion. In the domain of syntactic machine translation, complexity arises because there and too many target language word types. To manage this complexity, we translate into target language clusterings of increasing vocabulary size. This approach gives dramatic speed-ups while actually increasing final translation quality. Contact: Karen Livescu, TTI-C klivescu@tti-c.org 834-2549 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090317/d86a7807/attachment.htm From macglashan at tti-c.org Wed Mar 18 08:00:20 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Wed Mar 18 09:05:28 2009 Subject: [TTIC Colloquium] TTI-C Talk: Leonid Sigal, University of Toronto References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <21D3D5EE013647A4ABBBAEE7D614548D@jmacglDPLFYD1> REMINDER When: Thursday, March 19th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Leonid Sigal, University of Toronto Title: Physics-based Models and Priors for Human Motion Tracking Recovery and analysis of human pose and motion from video is the key enabling technology for a broad spectrum of applications, in and outside of computer science; including applications in HCI, biometrics, biomechanics and computer graphics. Despite years of research, the general problem of tracking a person in an unconstrained environment, particularly from monocular observations, remains challenging. In this talk I will describe the basic building blocks and challenges of the articulated human pose estimation and tracking, as well as my contributions to the various aspects of this problem and the field in recent years. I will particularly focus on the new and unique class of models that incorporate physic-based predictions and simulation into the inference process. Physics plays an important and intricate role in characterizing, describing and predicting human motion. The key benefit of using physics-based models or priors for tracking is the improved realism in the recovered motions, as well as enhanced ability to deal with weak image observations and diverse environmental interactions. Newtonian physics, in these models, approximates the rigid-body dynamics of the body in the environment through the application and integration of forces. Since the motion of the body is intimately tied with the environment, the use of such models also allows one to start reasoning about the geometry and physical properties of the environment as a whole (e.g. orientation and compliance of ground). This work is part of joint projects with colleagues at Brown University and University of Toronto. Bio: Leonid Sigal is a postdoctoral fellow in the Department of Computer Science at University of Toronto. He received his Ph.D. in computer science from Brown University (2007); his M.S. from Brown University (2003); his M.A. from Boston University (1999); and his B.Sc. degrees in Computer Science and Mathematics from Boston University (1999). From 1999 to 2001, he worked as a senior vision engineer at Cognex Corporation, where he developed industrial vision applications for pattern analysis and verification. In 2002, he spent a semester as a research intern at Siemens Corporate Research (SCR) working on autonomous obstacle detection and avoidance for vehicle navigation. During the summers of 2005 and 2006, he worked as a research intern at Intel Applications Research Lab (ARL) on human pose estimation and tracking. His work received the Best Paper Award at the Articulate Motion and Deformable Objects Conference in 2006 (with Prof. Michael J. Black). Dr. Sigal's research interests mainly lie in the areas of computer vision and machine learning, but also borderline fields of computer graphics, psychology and humanoid robotics. He is particularly interested in statistical models for problems of visual inference, including human motion recovery and analysis, graphical models, probabilistic and hierarchical inference. 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/20090318/a061a3d0/attachment-0001.htm From macglashan at tti-c.org Wed Mar 18 08:27:39 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Wed Mar 18 09:32:53 2009 Subject: [TTIC Colloquium] TTI-C Talk: Anastasios Sidiropoulos, University of Toronto References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: When: Tuesday, March 24th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Anastasios Sidiropoulos (University of Toronto) Title: Algorithmic metric embeddings We consider the problem of computing a low-distortion embedding of a finite metric space into low-dimensional and topologically simple spaces. It has been shown by Matousek [Mat90] that for any d\geq 1, any n-point metric can be embedded into R^d with distortion ~O(n^{2/d}) via a random projection, and that in the worst case this bound is essentially optimal. This clearly also implies an ~O(n^{2/d})-approximation algorithm for minimizing the distortion. We show that for any fixed d\geq 2, there is no polynomial-time algorithm for embedding into R^d, with approximation ratio better than Omega(n^{1/(22d)}), unless P=NP. Our result establishes that random projection is not too far from the best possible approximation algorithm for this problem. We also give some positive results by resorting to the relaxed notion of stochastic embeddings. Such mappings are allowed to be randomized, and the distance between every pair of points is required to be approximately preserved in expectation. We exhibit a novel approach for topological simplification of a metric space via stochastic embeddings. More precisely, we show how to stochastically embed bounded-genus graphs into planar graphs with constant distortion, and we discuss extensions to arbitrary minor-free graph families. This result is based on a structural theorem of Robertson and Seymour [RS83], and leads to a simplification of the Gupta-Newman-Rabinovich-Sinclair conjecture [GNRS04] which characterizes all graphs that have an O(1)-approximate multi-commodity max-flow/min-cut theorem. Contact: Julia Chuzhoy, TTI-C cjulia@tti-c.org 834-2572 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090318/f8870b10/attachment.htm From macglashan at tti-c.org Thu Mar 19 10:24:58 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Thu Mar 19 11:30:53 2009 Subject: [TTIC Colloquium] TTI-C Talk: Abhinav Gupta (University of Maryland, College Park) References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: When: Wednesday, March 25th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Abhinav Gupta (University of Maryland, College Park) Title: Beyond Nouns and Verbs: Learning Visually Grounded Stories of Images and Videos using Language and Vision In this talk, I will present our recent work on exploring synergy between language and vision for learning visually grounded contextual structures. Our work departs from the traditional view to visual and semantic learning where individual detectors and relationships are learned separately. Our work focuses on simultaneous learning of visual appearance and contextual models from richly annotated, weakly labeled datasets. In the first part of the talk, I will show how rich annotations can be utilized to constrain the learning of visually grounded models of nouns, prepositions and comparative adjectives from weakly labeled data. I will also show how visually grounded models of prepositions and comparative adjectives can be utilized as contextual models for scene analysis. In the second part, I will present storyline models for interpretation of videos. Storyline models go beyond pair-wise contextual models and represent higher order constraints by allowing only a few and finite number of possible action sequences (stories). Visual inference using storyline models involve inferring the "plot" of the video (sequence of actions) and recognizing individual activities in the plot. I will also present an iterative approach to learn visually grounded storyline models from video and linguistic information provided in captions. 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/20090319/495b8b5e/attachment.htm From macglashan at tti-c.org Mon Mar 23 08:06:38 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Mar 23 09:14:57 2009 Subject: [TTIC Colloquium] CANCELLED: TTI-C Talk: Abhinav Gupta (University of Maryland, College Park) References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: ***THIS TALK HAS BEEN CANCELLED. IT WILL BE RESCHEDULED AT A LATER DATE*** When: Wednesday, March 25th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Abhinav Gupta (University of Maryland, College Park) Title: Beyond Nouns and Verbs: Learning Visually Grounded Stories of Images and Videos using Language and Vision In this talk, I will present our recent work on exploring synergy between language and vision for learning visually grounded contextual structures. Our work departs from the traditional view to visual and semantic learning where individual detectors and relationships are learned separately. Our work focuses on simultaneous learning of visual appearance and contextual models from richly annotated, weakly labeled datasets. In the first part of the talk, I will show how rich annotations can be utilized to constrain the learning of visually grounded models of nouns, prepositions and comparative adjectives from weakly labeled data. I will also show how visually grounded models of prepositions and comparative adjectives can be utilized as contextual models for scene analysis. In the second part, I will present storyline models for interpretation of videos. Storyline models go beyond pair-wise contextual models and represent higher order constraints by allowing only a few and finite number of possible action sequences (stories). Visual inference using storyline models involve inferring the "plot" of the video (sequence of actions) and recognizing individual activities in the plot. I will also present an iterative approach to learn visually grounded storyline models from video and linguistic information provided in captions. 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/20090323/3a563a4f/attachment-0001.htm From macglashan at tti-c.org Mon Mar 23 10:40:43 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Mar 23 11:49:08 2009 Subject: [TTIC Colloquium] TTI-C Talk: Anastasios Sidiropoulos, University of Toronto References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: REMINDER When: Tuesday, March 24th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Anastasios Sidiropoulos (University of Toronto) Title: Algorithmic metric embeddings We consider the problem of computing a low-distortion embedding of a finite metric space into low-dimensional and topologically simple spaces. It has been shown by Matousek [Mat90] that for any d\geq 1, any n-point metric can be embedded into R^d with distortion ~O(n^{2/d}) via a random projection, and that in the worst case this bound is essentially optimal. This clearly also implies an ~O(n^{2/d})-approximation algorithm for minimizing the distortion. We show that for any fixed d\geq 2, there is no polynomial-time algorithm for embedding into R^d, with approximation ratio better than Omega(n^{1/(22d)}), unless P=NP. Our result establishes that random projection is not too far from the best possible approximation algorithm for this problem. We also give some positive results by resorting to the relaxed notion of stochastic embeddings. Such mappings are allowed to be randomized, and the distance between every pair of points is required to be approximately preserved in expectation. We exhibit a novel approach for topological simplification of a metric space via stochastic embeddings. More precisely, we show how to stochastically embed bounded-genus graphs into planar graphs with constant distortion, and we discuss extensions to arbitrary minor-free graph families. This result is based on a structural theorem of Robertson and Seymour [RS83], and leads to a simplification of the Gupta-Newman-Rabinovich-Sinclair conjecture [GNRS04] which characterizes all graphs that have an O(1)-approximate multi-commodity max-flow/min-cut theorem. Contact: Julia Chuzhoy, TTI-C cjulia@tti-c.org 834-2572 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090323/c92c4848/attachment.htm From macglashan at tti-c.org Thu Mar 26 08:56:57 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Thu Mar 26 09:58:09 2009 Subject: [TTIC Colloquium] TTI-C Talk: Abhinav Gupta (University of Maryland, College Park) References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <23594138D3084C20B469DEECD39AD5B5@jmacglDPLFYD1> NEW DATE When: Wednesday, April 1st @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Abhinav Gupta (University of Maryland, College Park) Title: Beyond Nouns and Verbs: Learning Visually Grounded Stories of Images and Videos using Language and Vision In this talk, I will present our recent work on exploring synergy between language and vision for learning visually grounded contextual structures. Our work departs from the traditional view to visual and semantic learning where individual detectors and relationships are learned separately. Our work focuses on simultaneous learning of visual appearance and contextual models from richly annotated, weakly labeled datasets. In the first part of the talk, I will show how rich annotations can be utilized to constrain the learning of visually grounded models of nouns, prepositions and comparative adjectives from weakly labeled data. I will also show how visually grounded models of prepositions and comparative adjectives can be utilized as contextual models for scene analysis. In the second part, I will present storyline models for interpretation of videos. Storyline models go beyond pair-wise contextual models and represent higher order constraints by allowing only a few and finite number of possible action sequences (stories). Visual inference using storyline models involve inferring the "plot" of the video (sequence of actions) and recognizing individual activities in the plot. I will also present an iterative approach to learn visually grounded storyline models from video and linguistic information provided in captions. 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/20090326/c59e0643/attachment.htm From macglashan at tti-c.org Tue Mar 31 08:12:23 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Mar 31 09:16:38 2009 Subject: [TTIC Colloquium] TTI-C Talk: Abhinav Gupta (University of Maryland, College Park) References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <56DD542C6B0940489267763158076C5F@jmacglDPLFYD1> REMINDER When: Wednesday, April 1st @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Abhinav Gupta (University of Maryland, College Park) Title: Beyond Nouns and Verbs: Learning Visually Grounded Stories of Images and Videos using Language and Vision In this talk, I will present our recent work on exploring synergy between language and vision for learning visually grounded contextual structures. Our work departs from the traditional view to visual and semantic learning where individual detectors and relationships are learned separately. Our work focuses on simultaneous learning of visual appearance and contextual models from richly annotated, weakly labeled datasets. In the first part of the talk, I will show how rich annotations can be utilized to constrain the learning of visually grounded models of nouns, prepositions and comparative adjectives from weakly labeled data. I will also show how visually grounded models of prepositions and comparative adjectives can be utilized as contextual models for scene analysis. In the second part, I will present storyline models for interpretation of videos. Storyline models go beyond pair-wise contextual models and represent higher order constraints by allowing only a few and finite number of possible action sequences (stories). Visual inference using storyline models involve inferring the "plot" of the video (sequence of actions) and recognizing individual activities in the plot. I will also present an iterative approach to learn visually grounded storyline models from video and linguistic information provided in captions. 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/20090331/e3eb93ca/attachment.htm