From macglashan at tti-c.org Mon Nov 2 08:16:15 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Nov 2 08:16:33 2009 Subject: [TTIC Colloquium] TTI-C Colloquium: S V N Vishwanathan, Purdue Message-ID: *REMINDER* When: *Monday, Nov 2 @ 1:00pm* Where: * TTI-C Conference Room #526*, 6045 S Kenwood Ave Who: *S V N Vishwanathan*, Purdue ( http://www.stat.purdue.edu/~vishy ) Title: * **A Quasi-Newton Approach to Nonsmooth Convex Optimization* Regularized risk minimization is at the heart of many machine learning algorithms. The underlying objective function to be minimized is convex, and often non-smooth. Classical optimization algorithms cannot handle this efficiently. In this talk we present our work on extending the well known BFGS quasi-Newton algorithm to handle non-smooth functions. Our extensions are justified both theoretically and experimentally. Joint work with Simon Guenter, Nic Schraudolph, Choon-Hui Teo and Jin Yu. Speaker Schedule: http://www.tti-c.org/colloquium.php Contact: Nati Srebro, TTI-C nati@tti-c.org 834-7493 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20091102/d3db42a9/attachment-0001.htm From macglashan at tti-c.org Tue Nov 3 11:20:45 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Nov 3 11:21:02 2009 Subject: [TTIC Colloquium] TTI-C Colloquium: Mathieu Salzmann, ICSI Berkeley Message-ID: When: *Monday, Nov 9 @ 1:00pm* Where: * TTI-C Conference Room #526*, 6045 S Kenwood Ave Who: *Mathieu Salzmann*, ICSI Berkeley Title: * **Reconstructing Deformable Surfaces: A Convex Formulation* In this talk, I will present an approach to reconstructing a deformable surface from a monocular input. I will first show that reconstruction from correspondences between an input image and a reference configuration can be formulated as the solution to a linear system, which leaves ambiguities in the recovered shape. To overcome these ambiguities, regularization terms must be introduced. I will therefore present distance inequality constraints that let us reconstruct folding surfaces. Not only are these constraints more general than state-of-the-art constraints, but they also yield a convex formulation of the reconstruction problem. Finally, to account for the lack of image information, I will introduce local deformation models. Such models give us the flexibility to regularize poorly-textured parts of the surface while not oversmoothing the well-textured ones. Bio: Since Febrary 2009, Mathieu Salzmann is a postdoctoral fellow in Prof. Trevor Darrell's group at the International Computer Science Institute and EECS at UC Berkeley. Previously, he completed a Ph.D with Prof. Pascal Fua at the Computer Vision laboratory at EPFL in Switzerland. He received his M.Sc. degree in Computer Science from EPFL in 2004. Contact: Raquel Urtasun, TTI-C rurtasun@tti-c.org 834-2550 Speaker Schedule: http://www.tti-c.org/colloquium.php -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20091103/f96a5e24/attachment.htm From macglashan at tti-c.org Mon Nov 9 09:38:36 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Nov 9 09:39:11 2009 Subject: [TTIC Colloquium] TTI-C Colloquium: Mathieu Salzmann, ICSI Berkeley Message-ID: *REMINDER* When: *Monday, Nov 9 @ 1:00pm* Where: * TTI-C Conference Room #526*, 6045 S Kenwood Ave Who: *Mathieu Salzmann*, ICSI Berkeley Title: * **Reconstructing Deformable Surfaces: A Convex Formulation* In this talk, I will present an approach to reconstructing a deformable surface from a monocular input. I will first show that reconstruction from correspondences between an input image and a reference configuration can be formulated as the solution to a linear system, which leaves ambiguities in the recovered shape. To overcome these ambiguities, regularization terms must be introduced. I will therefore present distance inequality constraints that let us reconstruct folding surfaces. Not only are these constraints more general than state-of-the-art constraints, but they also yield a convex formulation of the reconstruction problem. Finally, to account for the lack of image information, I will introduce local deformation models. Such models give us the flexibility to regularize poorly-textured parts of the surface while not oversmoothing the well-textured ones. Bio: Since Febrary 2009, Mathieu Salzmann is a postdoctoral fellow in Prof. Trevor Darrell's group at the International Computer Science Institute and EECS at UC Berkeley. Previously, he completed a Ph.D with Prof. Pascal Fua at the Computer Vision laboratory at EPFL in Switzerland. He received his M.Sc. degree in Computer Science from EPFL in 2004. Contact: Raquel Urtasun, TTI-C rurtasun@tti-c.org 834-2550 Speaker Schedule: http://www.tti-c.org/colloquium.php -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20091109/d1fc0b84/attachment.htm From macglashan at tti-c.org Tue Nov 10 09:37:37 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Nov 10 09:38:00 2009 Subject: [TTIC Colloquium] TTI-C Colloquium: Yali Amit, U of C Message-ID: When: *Monday, Nov 16 @ 1:00pm* Where: * TTI-C Conference Room #526*, 6045 S Kenwood Ave Who: *Yali Amit*, University of Chicago, CS Department Title: * **Statistical Models in Computer Vision* The goal of Computer Vision is the automatic labeling of images containing multiple objects as well as noise and clutter. Recent work has focused on two main tasks. The first is the classification among object classes in segmented images containing only one object and the second is the detection of a particular object class in a large image. Both tasks have been primarily addressed using discriminative learning. It is not clear however how these methods can extend to deal with the recognition of multiple object classes in images containing a number of objects in a wide range of configurations. I will present an approach which starts from simple statistical models for individual objects. With these models the important notion of invariance can be clearly formulated. Furthermore the individual object models can be composed to define models for object configurations. Decisions are likelihood based and do not depend on pretrained decision boundaries. I will briefly discuss some computational strategies for computing the scene annotation, show some applications, and describe some major difficulties we face in making further progress. Contact: David McAllester, TTI-C mcallester@tti-c.org 834-2550 Speaker Schedule: http://www.tti-c.org/colloquium.php -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20091110/e7fde4de/attachment.htm From macglashan at tti-c.org Mon Nov 16 08:27:06 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Nov 16 08:27:39 2009 Subject: [TTIC Colloquium] TTI-C Colloquium: Yali Amit, U of C Message-ID: *REMINDER* When: *Monday, Nov 16 @ 1:00pm* Where: * TTI-C Conference Room #526*, 6045 S Kenwood Ave Who: *Yali Amit*, University of Chicago, CS Department Title: * **Statistical Models in Computer Vision* The goal of Computer Vision is the automatic labeling of images containing multiple objects as well as noise and clutter. Recent work has focused on two main tasks. The first is the classification among object classes in segmented images containing only one object and the second is the detection of a particular object class in a large image. Both tasks have been primarily addressed using discriminative learning. It is not clear however how these methods can extend to deal with the recognition of multiple object classes in images containing a number of objects in a wide range of configurations. I will present an approach which starts from simple statistical models for individual objects. With these models the important notion of invariance can be clearly formulated. Furthermore the individual object models can be composed to define models for object configurations. Decisions are likelihood based and do not depend on pretrained decision boundaries. I will briefly discuss some computational strategies for computing the scene annotation, show some applications, and describe some major difficulties we face in making further progress. Contact: David McAllester, TTI-C mcallester@tti-c.org 834-2550 Speaker Schedule: http://www.tti-c.org/colloquium.php -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20091116/d2c7dd5a/attachment-0001.htm From macglashan at tti-c.org Tue Nov 17 12:03:02 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Nov 17 12:03:32 2009 Subject: [TTIC Colloquium] TTI-C Talk: Dhruv Batra, CMU Message-ID: When: *Tuesday, Nov 24 @ 10:30am* Where: * TTI-C Conference Room #526*, 6045 S Kenwood Ave Who: *Dhruv Batra*, CMU Title: * **Beyond Trees: MRF Inference via Outer-Planar Decomposition* A number of computer vision tasks can be formulated as discrete labelling problems. Markov Random Fields (MRFs) provide natural mathematical frameworks for modelling and solving these labelling problems. Maximum a posteriori (MAP) inference in MRFs (or energy minimization) is known to be NP-hard in general, and thus research has focussed on either finding efficiently solvable subclasses (such as trees and polytrees via Belief Propagation), or approximate inference algorithms (such as Loopy Belief Propagation and Tree-reweighted message passing). More recently, it has been shown that MAP inference on outer-planar graphs can be performed by formulating as a max-cut problem in planar graphs. In this work, we leverage this new class amenable to exact inference, and propose an approximate inference algorithm called Outer-Planar Decomposition (OPD). OPD involves ``decomposing'' an arbitrary energy function into energy functions over outer-planar graphs, and then a message passing algorithm over these outer-planar graphs. OPD is a strict generalization of tree-reweighted methods and contains as special cases each of the three TRW algorithms -- TRW-T, TRW-S and TRW-DD. Our Experiments show that OPD significantly outperforms current state of art inference methods -- TRW, QPBO and BP. Time permitting, I will also briefly talk about my work on interactive co-segmentation of related images -- present a novel distance metric algorithm and show a video about of our system presented at the CVPR '09 Demo session. Our recommendation system is able to guide users as to which image (and where) to scribble next. User studies showed our system is less cumbersome for users and facilitates quicker segmentation of many related images, as compared to a baseline set up. Joint work with Andrew Gallagher (Eastman Kodak) and Devi Parikh (TTI-C) (* Unpublished work) 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/20091117/38482e6f/attachment.htm From macglashan at tti-c.org Wed Nov 18 09:52:40 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Wed Nov 18 09:53:07 2009 Subject: [TTIC Colloquium] TTI-C Colloquium: Beata Beigman Klebanov, Northwestern Message-ID: When: *Monday, Nov 23 @ 1:00pm* Where: * TTI-C Conference Room #526*, 6045 S Kenwood Ave Who: *Beata Beigman Klebanov*, Northwestern University Title: * **Learning with Annotation Noise*** The success of supervised machine learning depends crucially on the quality of the labeled data. In computational linguistics, labeled data often come from human annotators analyzing texts for some latent content, like syntactic structures or the ideological perspective of the author. In such annotation projects, it is often assumed that the noise in the annotated data is random, hence not much work is devoted to identifying the noisy instances and assessing their impact on learning. Building on a case study of patterns of disagreements between human annotators in a binary classification task, we articulate an annotation generation model that allows a theoretical discussion of the type of noise introduced into the data through the annotation process. We show that this noise is different from both random classification noise and from malicious noise, and discuss its properties as far as learning is concerned. We then generalize the model of annotation generation and show that it fits well an existing dataset recently used in a benchmark. We conclude with a discussion of the implications of the findings for benchmarking practices in computational linguistics and elsewhere. Contact: Joseph Keshet, TTI-C jkeshet@tti-c.org 834-6850 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20091118/2ab97c0a/attachment.htm From macglashan at tti-c.org Mon Nov 23 09:48:17 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Nov 23 09:48:43 2009 Subject: [TTIC Colloquium] TTI-C Colloquium: Beata Beigman Klebanov, Northwestern Message-ID: When: *Monday, Nov 23 @ 1:00pm* Where: * TTI-C Conference Room #526*, 6045 S Kenwood Ave Who: *Beata Beigman Klebanov*, Northwestern University Title: * **Learning with Annotation Noise* The success of supervised machine learning depends crucially on the quality of the labeled data. In computational linguistics, labeled data often come from human annotators analyzing texts for some latent content, like syntactic structures or the ideological perspective of the author. In such annotation projects, it is often assumed that the noise in the annotated data is random, hence not much work is devoted to identifying the noisy instances and assessing their impact on learning. Building on a case study of patterns of disagreements between human annotators in a binary classification task, we articulate an annotation generation model that allows a theoretical discussion of the type of noise introduced into the data through the annotation process. We show that this noise is different from both random classification noise and from malicious noise, and discuss its properties as far as learning is concerned. We then generalize the model of annotation generation and show that it fits well an existing dataset recently used in a benchmark. We conclude with a discussion of the implications of the findings for benchmarking practices in computational linguistics and elsewhere. Contact: Joseph Keshet, TTI-C jkeshet@tti-c.org 834-6850 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20091123/ac1de010/attachment.htm From macglashan at tti-c.org Mon Nov 23 12:56:26 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Nov 23 12:57:06 2009 Subject: [TTIC Colloquium] TTI-C Talk: Dhruv Batra, CMU Message-ID: When: *Tuesday, Nov 24 @ 10:30am* Where: * TTI-C Conference Room #526*, 6045 S Kenwood Ave Who: *Dhruv Batra*, CMU Title: * **Beyond Trees: MRF Inference via Outer-Planar Decomposition* A number of computer vision tasks can be formulated as discrete labelling problems. Markov Random Fields (MRFs) provide natural mathematical frameworks for modelling and solving these labelling problems. Maximum a posteriori (MAP) inference in MRFs (or energy minimization) is known to be NP-hard in general, and thus research has focussed on either finding efficiently solvable subclasses (such as trees and polytrees via Belief Propagation), or approximate inference algorithms (such as Loopy Belief Propagation and Tree-reweighted message passing). More recently, it has been shown that MAP inference on outer-planar graphs can be performed by formulating as a max-cut problem in planar graphs. In this work, we leverage this new class amenable to exact inference, and propose an approximate inference algorithm called Outer-Planar Decomposition (OPD). OPD involves ``decomposing'' an arbitrary energy function into energy functions over outer-planar graphs, and then a message passing algorithm over these outer-planar graphs. OPD is a strict generalization of tree-reweighted methods and contains as special cases each of the three TRW algorithms -- TRW-T, TRW-S and TRW-DD. Our Experiments show that OPD significantly outperforms current state of art inference methods -- TRW, QPBO and BP. Time permitting, I will also briefly talk about my work on interactive co-segmentation of related images -- present a novel distance metric algorithm and show a video about of our system presented at the CVPR '09 Demo session. Our recommendation system is able to guide users as to which image (and where) to scribble next. User studies showed our system is less cumbersome for users and facilitates quicker segmentation of many related images, as compared to a baseline set up. Joint work with Andrew Gallagher (Eastman Kodak) and Devi Parikh (TTI-C) (* Unpublished work) 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/20091123/915d9e9f/attachment-0001.htm