From macglashan at tti-c.org Thu Apr 2 08:56:16 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Thu Apr 2 09:56:20 2009 Subject: [TTIC Colloquium] TTI-C Talk: Andreas Argyriou, University College London References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <917518B02F17476E90C36A95F9B71DBF@jmacglDPLFYD1> When: Tuesday, April 7th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Andreas Argyriou, University College London Title: Multi-Task Learning and Matrix Regularization Multi-task learning extends the standard paradigm of supervised learning. In multi-task learning, samples for multiple related tasks are given and the goal is to learn a function for each task and also to generalize well (transfer learned knowledge) on new tasks. The applications of this paradigm are numerous and range from computer vision to collaborative filtering to bioinformatics while it also relates to vector valued problems, multiclass, multiview learning etc. I will present a framework for multi-task learning which is based on learning a common kernel for all tasks. I will also show how this formulation connects to the trace norm and group Lasso approaches. Moreover, the proposed optimization problem can be solved using an alternating minimization algorithm which is simple and efficient. It can also be "kernelized" by virtue of a multi-task representer theorem, which holds for a large family of matrix regularization problems and includes the classical representer theorem as a special case. Finally, I will draw an analogy between multi-task learning and convex kernel learning and will present a general convergent algorithm for learning convex combinations of finite or infinite kernels. 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/20090402/07d47c35/attachment-0001.htm From macglashan at tti-c.org Fri Apr 3 13:20:06 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Fri Apr 3 14:20:55 2009 Subject: [TTIC Colloquium] TTI-C Talk: Nate Foster, UPenn References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <290D132A79124B32A794CFBC0CF7DF93@jmacglDPLFYD1> When: Friday, April 10th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Nate Foster, UPenn Title: Bidirectional Programming Languages The need to edit data through a view arises in a host of applications across many different areas of computing. Unfortunately, few existing systems have support for updatable views and so, in practice, they are usually implemented using two separate programs -- one to compute views from sources, and another to handle updates. This rudimentary design is tedious to program, difficult to reason about, and a nightmare to maintain. In this talk, I will present bidirectional programming languages, which provide an elegant and effective mechanism for describing updatable views. Unlike programs written in an ordinary language, which only work in one direction, programs in a bidirectional language can be run both forwards and backwards. When read from left to right, they describe functions that map sources to views. When read from right to left, they describe functions that map updated views back to updated sources. Besides eliminating redundancy, these languages can be designed to ensure correctness, guaranteeing by construction that the two transformations work well together. Starting from the foundations, I will describe a general semantic space of well-behaved bidirectional transformations called lenses. Then, building on this framework, I will describe a specific language for writing lenses on strings, with a syntax and type system based on the familiar regular operators (union, concatenation, and Kleene star). Finally, I will describe a collection of extensions addressing the subtle complications that arise when lenses are used to manipulate ordered, ignorable, and confidential data. BIO: Nate Foster will receive his PhD in Computer and Information Science from the University of Pennsylvania in the summer of 2009. His research centers around problems in programming languages, data management, and security. Before coming to Penn, he studied at the University of Cambridge (MPhil in History and Philosophy of Science) and Williams College (BA in Computer Science). He has also worked as an intern at INRIA Rhone-Alpes and IBM Research. His dissertation was supported by an NSF Graduate Research Fellowship and was recently selected as a winner of Penn's Rubinoff Award. 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/20090403/5ec88b06/attachment.htm From cnovak at tti-c.org Mon Apr 6 09:32:09 2009 From: cnovak at tti-c.org (Chrissy Novak) Date: Mon Apr 6 10:39:33 2009 Subject: [TTIC Colloquium] TTI-C Talk: Andreas Argyriou, University College London Message-ID: <528B2075404A4234A2FC69BEFA11523E@cnovakHBRQFD1> REMINDER When: Tuesday, April 7th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Andreas Argyriou, University College London Title: Multi-Task Learning and Matrix Regularization Multi-task learning extends the standard paradigm of supervised learning. In multi-task learning, samples for multiple related tasks are given and the goal is to learn a function for each task and also to generalize well (transfer learned knowledge) on new tasks. The applications of this paradigm are numerous and range from computer vision to collaborative filtering to bioinformatics while it also relates to vector valued problems, multiclass, multiview learning etc. I will present a framework for multi-task learning which is based on learning a common kernel for all tasks. I will also show how this formulation connects to the trace norm and group Lasso approaches. Moreover, the proposed optimization problem can be solved using an alternating minimization algorithm which is simple and efficient. It can also be "kernelized" by virtue of a multi-task representer theorem, which holds for a large family of matrix regularization problems and includes the classical representer theorem as a special case. Finally, I will draw an analogy between multi-task learning and convex kernel learning and will present a general convergent algorithm for learning convex combinations of finite or infinite kernels. 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/20090406/67c92890/attachment.htm From macglashan at tti-c.org Tue Apr 7 08:53:15 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Apr 7 09:56:23 2009 Subject: [TTIC Colloquium] TTI-C Talk: Neelakantan Krishnaswami, CMU References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: When: Thursday, April 9th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Neelakantan Krishnaswami, CMU Title: Proving GUIs Correct: Verifying Higher-Order Imperative Programs with Higher-Order Separation Logic O'Hearn and Reynolds' separation logic has proven to be a very successful attempt at taming many of the difficulties associated with reasoning about aliased, mutable data structures. Using it, researchers have given correctness proofs of even quite intricate low-level imperative programs such as garbage collectors and device drivers. However, high level languages such as ML and Haskell also give programmers access to mutable, aliased data, and when those features are used, programmers are still prone to all the troubles state is heir to. In fact, many problems become more complex, since these languages encourage the use of an abstract, higher-order style, and support the design of libraries that rely on higher-order functions as well as callbacks (ie, references to functions in the heap). In this talk, I'll describe work I've done (in collaboration with my PhD supervisors) designing a version of separation logic suitable for use in languages such as ML, and describe an application of this logic to formally verifying the correctness of a small library for writing event-driven programs in a lazy dataflow style. This then allows an efficient imperative implementation of a functional reactive programming library. 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/20090407/21d4c74a/attachment-0001.htm From macglashan at tti-c.org Tue Apr 7 15:48:37 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Apr 7 16:51:53 2009 Subject: [TTIC Colloquium] Math Dept- Antoni Zygmund and Alberto P. Calderon Lectures Message-ID: <2673F5A2125347F5802D13876508CE35@jmacglDPLFYD1> The University of Chicago DEPARTMENT OF MATHEMATICS announces the 2009 Antoni Zygmund and Alberto P. Calderon Lectures by Maciej Zworski University of California/Berkeley THREE LECTURES I. What is Microlocal Analysis? Monday, April 20, 4:00 - Room 202 Eckhart Hall - 1118 E. 58th Street II. What are Quantum Resonances? Tuesday, April 21, 4:30 - Room 206 Eckhart Hall - 1118 E. 58th Street III. Microlocal Methods in the Study of Resonances Wednesday, April 22, 4:00 - Room 203 Eckhart Hall - 1118 E. 58th Street Tea will be served in the Common Room, Eckhart 209, 30 minutes before each lecture. ___________________________________________________________________________ Persons with a disability who believe they may need assistance please call Lynette Whalum at (773) 702-7100. From macglashan at tti-c.org Wed Apr 8 14:27:33 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Wed Apr 8 14:27:46 2009 Subject: [TTIC Colloquium] TTI-C Talk: Neelakantan Krishnaswami, CMU References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <707DAB07854941E3AE2CB73AD1DAC857@jmacglDPLFYD1> REMINDER When: Thursday, April 9th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Neelakantan Krishnaswami, CMU Title: Proving GUIs Correct: Verifying Higher-Order Imperative Programs with Higher-Order Separation Logic O'Hearn and Reynolds' separation logic has proven to be a very successful attempt at taming many of the difficulties associated with reasoning about aliased, mutable data structures. Using it, researchers have given correctness proofs of even quite intricate low-level imperative programs such as garbage collectors and device drivers. However, high level languages such as ML and Haskell also give programmers access to mutable, aliased data, and when those features are used, programmers are still prone to all the troubles state is heir to. In fact, many problems become more complex, since these languages encourage the use of an abstract, higher-order style, and support the design of libraries that rely on higher-order functions as well as callbacks (ie, references to functions in the heap). In this talk, I'll describe work I've done (in collaboration with my PhD supervisors) designing a version of separation logic suitable for use in languages such as ML, and describe an application of this logic to formally verifying the correctness of a small library for writing event-driven programs in a lazy dataflow style. This then allows an efficient imperative implementation of a functional reactive programming library. 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/20090408/d7d7f475/attachment.htm From macglashan at tti-c.org Thu Apr 9 09:44:34 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Thu Apr 9 09:44:40 2009 Subject: [TTIC Colloquium] TTI-C Talk: Nate Foster, UPenn References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: REMINDER When: Friday, April 10th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Nate Foster, UPenn Title: Bidirectional Programming Languages The need to edit data through a view arises in a host of applications across many different areas of computing. Unfortunately, few existing systems have support for updatable views and so, in practice, they are usually implemented using two separate programs -- one to compute views from sources, and another to handle updates. This rudimentary design is tedious to program, difficult to reason about, and a nightmare to maintain. In this talk, I will present bidirectional programming languages, which provide an elegant and effective mechanism for describing updatable views. Unlike programs written in an ordinary language, which only work in one direction, programs in a bidirectional language can be run both forwards and backwards. When read from left to right, they describe functions that map sources to views. When read from right to left, they describe functions that map updated views back to updated sources. Besides eliminating redundancy, these languages can be designed to ensure correctness, guaranteeing by construction that the two transformations work well together. Starting from the foundations, I will describe a general semantic space of well-behaved bidirectional transformations called lenses. Then, building on this framework, I will describe a specific language for writing lenses on strings, with a syntax and type system based on the familiar regular operators (union, concatenation, and Kleene star). Finally, I will describe a collection of extensions addressing the subtle complications that arise when lenses are used to manipulate ordered, ignorable, and confidential data. BIO: Nate Foster will receive his PhD in Computer and Information Science from the University of Pennsylvania in the summer of 2009. His research centers around problems in programming languages, data management, and security. Before coming to Penn, he studied at the University of Cambridge (MPhil in History and Philosophy of Science) and Williams College (BA in Computer Science). He has also worked as an intern at INRIA Rhone-Alpes and IBM Research. His dissertation was supported by an NSF Graduate Research Fellowship and was recently selected as a winner of Penn's Rubinoff Award. 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/20090409/cb6f6b81/attachment.htm From macglashan at tti-c.org Fri Apr 10 13:57:30 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Fri Apr 10 13:57:45 2009 Subject: [TTIC Colloquium] TTI-C Talk: Pradeep Ravikumar, UC Berkeley References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <569F59CC268F4E3AB67C5786D9868083@jmacglDPLFYD1> When: Tuesday, April 14th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Pradeep Ravikumar, UC Berkeley Title: Sparse Model Estimation: Parametric and Nonparametric Settings A common approach in settings with high-dimensional data has been to estimate models that are ``sparse,'' in the sense that an index set of relevant model components has small cardinality. In this talk I will cover two instances, one parametric and the other nonparametric, of sparse model estimation. The first part of the talk considers the task of estimating the covariance and inverse covariance or concentration matrices of a random vector from i.i.d. observations. We study an estimator based on minimizing an l1-penalized log-determinant Bregman divergence, that is equivalent to the usual l1-regularized maximum likelihood estimator when the random vector is multivariate Gaussian. We analyze the performance of this estimator under high-dimensional scaling, in which the number of variables and other model parameters are allowed to grow as a function of the sample size. Our analysis identifies key players affecting the convergence rates of the estimator in various norms as well as its success in recovering the true sparsity pattern (its ``sparsistency''). The second part of the talk considers the task of encoding fMRI signals from the primary visual cortex, also called area V1, of the brain in response to natural image stimuli; as well as identifying potential features of images that drive the neural activity. Our method is based on the understanding that the fMRI signal reflects the pooled, and potentially nonlinearly transformed output of a large population of neurons in area V1. Our class of models, which we call the V-SPAM framework, mimics this with an initial hierarchical filtering stage that consists of three layers of artificial neuronal cells, and a final nonparametric pooling stage which learns nonparametric transformations of a sparse set of neuronal filters. This is joint work with Garvesh Raskutti, Vincent Vu, Martin Wainwright, Bin Yu, and the Jack Gallant lab at UC Berkeley; Kendrick Kay, Thomas Naselaris and Jack Gallant. 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/20090410/1f42df8c/attachment-0001.htm From macglashan at tti-c.org Mon Apr 13 09:14:10 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Apr 13 09:14:22 2009 Subject: [TTIC Colloquium] TTI-C Talk: Pradeep Ravikumar, UC Berkeley References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <28F7D40BBF1B49EBABCAAB2ECF3F7326@jmacglDPLFYD1> REMINDER When: Tuesday, April 14th @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Pradeep Ravikumar, UC Berkeley Title: Sparse Model Estimation: Parametric and Nonparametric Settings A common approach in settings with high-dimensional data has been to estimate models that are ``sparse,'' in the sense that an index set of relevant model components has small cardinality. In this talk I will cover two instances, one parametric and the other nonparametric, of sparse model estimation. The first part of the talk considers the task of estimating the covariance and inverse covariance or concentration matrices of a random vector from i.i.d. observations. We study an estimator based on minimizing an l1-penalized log-determinant Bregman divergence, that is equivalent to the usual l1-regularized maximum likelihood estimator when the random vector is multivariate Gaussian. We analyze the performance of this estimator under high-dimensional scaling, in which the number of variables and other model parameters are allowed to grow as a function of the sample size. Our analysis identifies key players affecting the convergence rates of the estimator in various norms as well as its success in recovering the true sparsity pattern (its ``sparsistency''). The second part of the talk considers the task of encoding fMRI signals from the primary visual cortex, also called area V1, of the brain in response to natural image stimuli; as well as identifying potential features of images that drive the neural activity. Our method is based on the understanding that the fMRI signal reflects the pooled, and potentially nonlinearly transformed output of a large population of neurons in area V1. Our class of models, which we call the V-SPAM framework, mimics this with an initial hierarchical filtering stage that consists of three layers of artificial neuronal cells, and a final nonparametric pooling stage which learns nonparametric transformations of a sparse set of neuronal filters. This is joint work with Garvesh Raskutti, Vincent Vu, Martin Wainwright, Bin Yu, and the Jack Gallant lab at UC Berkeley; Kendrick Kay, Thomas Naselaris and Jack Gallant. 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/20090413/62439448/attachment.htm From macglashan at tti-c.org Tue Apr 14 08:34:12 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Tue Apr 14 08:34:28 2009 Subject: [TTIC Colloquium] TTI-C Talk: Devi Parikh, CMU References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <6A37498BA649418FBE4DD5FAD477A4AC@jmacglDPLFYD1> When: Tuesday, April 21st @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Devi Parikh, Carnegie Mellon University Title: The Role of Context in Image Understanding: When, For What, and How? A key problem in computer vision is image understanding, which we define as the task of recognizing objects in the scene, and perhaps the scene category itself. Traditionally, object recognition has been accomplished by considering only the information within the object to be recognized. Incorporating contextual information, i.e., information outside the boundaries of the object, for enhanced recognition has received significant attention in recent works. In this talk, we take a closer look at the role of context. Specifically, we ask three questions. First: When is context really helpful? We show, through computer vision experiments as well as human studies, that context provides improvements in recognition performances only when the appearance information is weak (such as in low resolution images or in the presence of occlusion). Second: For what tasks can contextual information be leveraged? We show that apart from high- level tasks of object recognition and detection, contextual information can be effectively leveraged for low level tasks as well, such as identifying salient or representative patches in an image. Lastly (the focus of the talk), How can context be learnt? Or alternatively, how much contextual information can be extracted in an unsupervised manner? We propose a unified hierarchical representation for contextual interactions or spatial patterns among visual entities at all levels, from low-level features to parts of objects, objects, groups of objects and ultimately the entire scene. We present results of our approach on a variety of datasets such as object categories, street scenes and natural scene images. Contact: Greg Shakhnarovich, TTI-C greg@tti-c.org 834-7493 -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090414/02b5ea39/attachment.htm From macglashan at tti-c.org Fri Apr 17 10:41:52 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Fri Apr 17 10:42:02 2009 Subject: [TTIC Colloquium] TTI-C Talk: Raquel Urtasun, UC Berkeley References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <7C77F78AC92043839F96E1065AFD24D7@jmacglDPLFYD1> When: Thursday, April 23rd @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Raquel Urtasun, UC Berkeley Title: Non-parametric models for the analysis of human behavior Understanding human behavior is an important research endeavor because its success can have a tremendous positive impact on our everyday lives. Health, biology, psychology, robotics and the game industry are some examples amongst the wide range of applications. The development of the web together with the reduction of storage and processing cost have recently made available a large amount of data key for the analysis of human behavior. This data can have very different properties and the underlying generative process can be unknown or very difficult to model. Non-parametric models are well suited for the analysis of human behavior since they make very few assumptions and the structure of the problem is determined from the data. In this talk I will show how Gaussian Processes can be used to model different aspects of human behavior including their motions, the world they live in, and their preferences. When dealing with high-dimensional data, it is desirable to reduce the dimensionality of the data while preserving the original information in the data distribution, allowing for more efficient learning and inference. Linear dimensionality reduction techniques and graph-based methods are popular approaches but can result in poor approximations when dealing with complex datasets or when the manifold assumption is violated, e.g., sparse noisy data. Non-linear latent variable models can recover complex manifolds but they suffer from local minima, since they rely on the optimization of complex non-linear functions that are non-convex. Moreover, there is no principled way to choose the dimensionality of the latent space. The rank of a matrix is often an efficient way to describe the complexity or dimensionality of a system. When learning low dimensional representations we would like to encourage the rank of the latent space to be small. In this talk, I will construct a relaxation of the rank minimization problem and build a prior over the latent space that encourages sparsity of the singular values resulting in low-dimensional representations. In doing so, our method is able to simultaneously estimate the latent space and its dimensionality in a continuous fashion. Gaussian processes scale poorly with the size of the training data, its computational complexity being O(N^3), with N the number of examples. In applications such as the prediction of user preferences the amount of data can be arbitrarily large, e.g., millions of examples for Netflix. In this talk I will show how to exploit the inherent sparseness of the data to design a stochastic gradient descent algorithm that can effectively learn non-linear representations from very large databases. Bio: Raquel Urtasun is currently a Postdoctoral Research Scientist at UC Berkeley EECS & ICSI working with Prof. Trevor Darrell. Her main research areas are computer vision, machine learning and computer graphics. During 2006-2008, she was a postdoctoral associate at MIT-CSAIL. She earned her PhD at EPFL (Switzerland) in 2006 under the supervision of Prof. Pascal Fua on Motion Models for Robust 3D Human Body Tracking. 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/20090417/cdaf8605/attachment-0001.htm From macglashan at tti-c.org Fri Apr 17 15:39:09 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Fri Apr 17 15:39:19 2009 Subject: [TTIC Colloquium] TTI-C Talk: Percy Liang, UC Berkeley References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: When: Wednesday, April 22nd @ 11:00am Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Percy Liang, UC Berkeley Title: Asymptotically Optimal Regularization It is well-known that regularization is crucial for good performance, and in recent years, machine learning has given rise to a diverse array of regularizers, especially in semi-supervised learning and multi-task learning. Which regularizer should one choose? In this talk, we present a general method for deriving the asymptotically optimal regularizer for a given loss function, which provides both insight and quantitative guidance. Joint work with Francis Bach, Guillaume Bouchard, and Michael Jordan Bio: Percy Liang is a fourth-year Ph.D. student at UC Berkeley working with Michael Jordan and Dan Klein. He graduated with a bachelors in computer science and math from MIT. He works actively in both natural language processing and machine learning, focusing on unsupervised learning of rich latent-variable probabilistic models and also developing theoretical analyses. He holds NDSEG and NSF fellowships and has received a best student paper award at ICML in 2008. He is also an aspiring concert pianist. 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/20090417/84a6b59d/attachment.htm From macglashan at tti-c.org Mon Apr 20 09:09:01 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Apr 20 09:09:14 2009 Subject: [TTIC Colloquium] TTI-C Talk: Devi Parikh, CMU References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <29840BF4AA97488A94D6A3A2F27958D5@jmacglDPLFYD1> REMINDER When: Tuesday, April 21st @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Devi Parikh, Carnegie Mellon University Title: The Role of Context in Image Understanding: When, For What, and How? A key problem in computer vision is image understanding, which we define as the task of recognizing objects in the scene, and perhaps the scene category itself. Traditionally, object recognition has been accomplished by considering only the information within the object to be recognized. Incorporating contextual information, i.e., information outside the boundaries of the object, for enhanced recognition has received significant attention in recent works. In this talk, we take a closer look at the role of context. Specifically, we ask three questions. First: When is context really helpful? We show, through computer vision experiments as well as human studies, that context provides improvements in recognition performances only when the appearance information is weak (such as in low resolution images or in the presence of occlusion). Second: For what tasks can contextual information be leveraged? We show that apart from high- level tasks of object recognition and detection, contextual information can be effectively leveraged for low level tasks as well, such as identifying salient or representative patches in an image. Lastly (the focus of the talk), How can context be learnt? Or alternatively, how much contextual information can be extracted in an unsupervised manner? We propose a unified hierarchical representation for contextual interactions or spatial patterns among visual entities at all levels, from low-level features to parts of objects, objects, groups of objects and ultimately the entire scene. We present results of our approach on a variety of datasets such as object categories, street scenes and natural scene images. 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/20090420/10b0901b/attachment.htm From macglashan at tti-c.org Wed Apr 22 08:17:29 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Wed Apr 22 08:17:39 2009 Subject: [TTIC Colloquium] TTI-C Talk: Percy Liang, UC Berkeley References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <3FB356BDAE1E4861AAC422388DB03E15@jmacglDPLFYD1> REMINDER When: TODAY, Wednesday, April 22nd @ 11:00am Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Percy Liang, UC Berkeley Title: Asymptotically Optimal Regularization It is well-known that regularization is crucial for good performance, and in recent years, machine learning has given rise to a diverse array of regularizers, especially in semi-supervised learning and multi-task learning. Which regularizer should one choose? In this talk, we present a general method for deriving the asymptotically optimal regularizer for a given loss function, which provides both insight and quantitative guidance. Joint work with Francis Bach, Guillaume Bouchard, and Michael Jordan Bio: Percy Liang is a fourth-year Ph.D. student at UC Berkeley working with Michael Jordan and Dan Klein. He graduated with a bachelors in computer science and math from MIT. He works actively in both natural language processing and machine learning, focusing on unsupervised learning of rich latent-variable probabilistic models and also developing theoretical analyses. He holds NDSEG and NSF fellowships and has received a best student paper award at ICML in 2008. He is also an aspiring concert pianist. 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/20090422/7061dedf/attachment.htm From macglashan at tti-c.org Wed Apr 22 11:23:23 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Wed Apr 22 11:23:31 2009 Subject: [TTIC Colloquium] TTI-C Talk: Raquel Urtasun, UC Berkeley References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <631D1855912444AC9870D4953DC3A0B3@jmacglDPLFYD1> REMINDER When: Thursday, April 23rd @ 11:00am (lunch will be provided after talk) Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th Floor) Who: Raquel Urtasun, UC Berkeley Title: Non-parametric models for the analysis of human behavior Understanding human behavior is an important research endeavor because its success can have a tremendous positive impact on our everyday lives. Health, biology, psychology, robotics and the game industry are some examples amongst the wide range of applications. The development of the web together with the reduction of storage and processing cost have recently made available a large amount of data key for the analysis of human behavior. This data can have very different properties and the underlying generative process can be unknown or very difficult to model. Non-parametric models are well suited for the analysis of human behavior since they make very few assumptions and the structure of the problem is determined from the data. In this talk I will show how Gaussian Processes can be used to model different aspects of human behavior including their motions, the world they live in, and their preferences. When dealing with high-dimensional data, it is desirable to reduce the dimensionality of the data while preserving the original information in the data distribution, allowing for more efficient learning and inference. Linear dimensionality reduction techniques and graph-based methods are popular approaches but can result in poor approximations when dealing with complex datasets or when the manifold assumption is violated, e.g., sparse noisy data. Non-linear latent variable models can recover complex manifolds but they suffer from local minima, since they rely on the optimization of complex non-linear functions that are non-convex. Moreover, there is no principled way to choose the dimensionality of the latent space. The rank of a matrix is often an efficient way to describe the complexity or dimensionality of a system. When learning low dimensional representations we would like to encourage the rank of the latent space to be small. In this talk, I will construct a relaxation of the rank minimization problem and build a prior over the latent space that encourages sparsity of the singular values resulting in low-dimensional representations. In doing so, our method is able to simultaneously estimate the latent space and its dimensionality in a continuous fashion. Gaussian processes scale poorly with the size of the training data, its computational complexity being O(N^3), with N the number of examples. In applications such as the prediction of user preferences the amount of data can be arbitrarily large, e.g., millions of examples for Netflix. In this talk I will show how to exploit the inherent sparseness of the data to design a stochastic gradient descent algorithm that can effectively learn non-linear representations from very large databases. Bio: Raquel Urtasun is currently a Postdoctoral Research Scientist at UC Berkeley EECS & ICSI working with Prof. Trevor Darrell. Her main research areas are computer vision, machine learning and computer graphics. During 2006-2008, she was a postdoctoral associate at MIT-CSAIL. She earned her PhD at EPFL (Switzerland) in 2006 under the supervision of Prof. Pascal Fua on Motion Models for Robust 3D Human Body Tracking. 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/20090422/d37db42d/attachment-0001.htm From macglashan at tti-c.org Thu Apr 23 10:34:06 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Thu Apr 23 10:34:19 2009 Subject: [TTIC Colloquium] TTI-C Talk: Ashutosh Saxena, Stanford References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: <69DEBAF8A2CA4A8A8309BE7F40F8AFBC@jmacglDPLFYD1> > When: Thursday, April 30th @ 11:00am (lunch will be provided > after talk) > > Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th > Floor) > > Who: Ashutosh Saxena, Stanford University > > Title: Robotic Grasping and Depth Perception: Learning 3D > Models from a Single Image > > The ability to perceive the 3D shape of the environment is a basic ability for a robot. We present an algorithm to convert standard digital pictures into 3D models. This is a challenging problem, since an image is formed by a projection of the 3D scene onto two dimensions, thus losing the depth information. We take a supervised learning approach to this problem, and use a Markov Random Field (MRF) to model the scene depth as a function of the image features. We show that, even on unstructured scenes of a large variety of environments, our algorithm is frequently able to recover accurate 3D models. We then apply our methods to robotics applications: (a) obstacle avoidance for autonomously driving a small electric car, and (b) robot manipulation, where we develop vision-based learning algorithms for grasping novel objects. This enables our robot to perform tasks such as open new doors, clear up cluttered tables, and unload items from a dishwasher. > 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/20090423/8d0f6797/attachment.htm From macglashan at tti-c.org Mon Apr 27 09:18:41 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Mon Apr 27 09:19:00 2009 Subject: [TTIC Colloquium] UoC Talk by David Xiao, Princeton University on May 11, 2009 Message-ID: DEPARTMENT OF COMPUTER SCIENCE UNIVERSITY OF CHICAGO Date: Monday, May 11, 2009 Time: 3:45 p.m. Place: RY 251 ---------------------------------------------------------- Speaker: David Xiao From: Princeton University Website: http://www.cs.princeton.edu/~dxiao/ Title: On the Black-box Complexity of PAC Learning Abstract: The PAC model (Valiant, CACM '84) is one of the central models studied in computational learning theory. There is evidence that many specific classes of functions (e.g. intersections of half- spaces, parity functions with noise, etc.) are hard to learn by efficient algorithms, and cryptographic assumptions imply that learning small circuits is hard. We say that PAC learning is hard if no efficient algorithm can learn all functions computable by small circuits. In this talk, we show that the black-box complexity of PAC learning lies strictly between NP and ZK. More precisely, if P = NP then PAC learning is easy and if ZK !=BPP then PAC learning is hard, but black- box techniques (with some additional restrictions) do not suffice to prove equivalence in either case. With regard to NP, we rule out non-adaptive reductions using a PAC learning oracle to solve an NP-complete problem by showing this would imply that coNP in contained in AM, which is considered implausible. With regard to ZK, we rule out relativizing proofs that ZK !=BPP based on hardness of learning. Using the characterization of ZK of Ong and Vadhan (EUROCRYPT '07), we also show that any black-box construction of a (computational) ZK protocol for a language L based on hardness of learning implies that L actually has a statistical zero knowledge proof (i.e. L is in SZK), and hence such a black-box construction is unlikely to exist for NP-complete languages. Parts of this talk are based on joint work with Benny Applebaum and Boaz Barak. Refreshments will be served prior to the talk in RY 255. _______________________________________________ Colloquium mailing list - Colloquium@mailman.cs.uchicago.edu https://mailman.cs.uchicago.edu/mailman/listinfo/colloquium -------------- next part -------------- An HTML attachment was scrubbed... URL: http://ttic.uchicago.edu/pipermail/colloquium/attachments/20090427/aa53cf21/attachment.htm From macglashan at tti-c.org Wed Apr 29 09:40:48 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Wed Apr 29 09:40:56 2009 Subject: [TTIC Colloquium] TTI-C Talk: Ashutosh Saxena, Stanford References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: REMINDER > When: Thursday, April 30th @ 11:00am (lunch will be provided > after talk) > > Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th > Floor) > > Who: Ashutosh Saxena, Stanford University > > Title: Robotic Grasping and Depth Perception: Learning 3D > Models from a Single Image > > > The ability to perceive the 3D shape of the environment is a basic ability > for a robot. We present an algorithm to convert standard digital pictures > into 3D models. > > This is a challenging problem, since an image is formed by a projection of > the 3D scene onto two dimensions, thus losing the depth information. > We take a supervised learning approach to this problem, and use a Markov > Random Field (MRF) to model the scene depth as a function of the image > features. We show that, even on unstructured scenes of a large variety of > environments, our algorithm is frequently able to recover accurate 3D > models. > > We then apply our methods to robotics applications: (a) obstacle avoidance > for autonomously driving a small electric car, and (b) robot manipulation, > where we develop vision-based learning algorithms for grasping novel > objects. This enables our robot to perform tasks such as open new doors, > clear up cluttered tables, and unload items from a dishwasher. > > 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/20090429/d0267006/attachment.htm From macglashan at tti-c.org Thu Apr 30 10:14:26 2009 From: macglashan at tti-c.org (Julia MacGlashan) Date: Thu Apr 30 10:14:37 2009 Subject: [TTIC Colloquium] TTI-C Talk: Xuefeng Zhou, Washington University in St Louis References: <9BFA4FFE1ACE407581765CA2326E1547@jmacglDPLFYD1> Message-ID: > When: Monday, May 4 @ 11:00am (lunch will be provided after > talk) > > Where: 6045 S Kenwood Ave, TTI-C Conference Room #526 (5th > Floor) > > Who: Xuefeng Zhou (Washington University in St Louis) > > Title: Study of microRNAs: a Biology Problem with > Computational Challenges > > Repression of gene expression is an important regulatory mechanism that controls many biological processes such as development, cell proliferation and differentiation. The discovery of microRNAs (miRNAs) has broadened our perspectives on the mechanisms of down-regulation of gene expression and shed light on an entirely novel level of post-transcriptional regulation. Besides their important functions in the development of animals and plants, miRNAs have been shown to play crucial roles in the pathogenesis of many diseases, such as cancer. Since the discovery of the very first miRNA, almost all progresses on study of miRNA resorted to the help from computational approaches. In this talk, I will first present our recent work on the prediction of novel miRNAs. Available computational methods rely on non-trivial number of known miRNAs, clear genome annotations and evolutionary conservation information. We developed a novel ranking algorithm based on random works to computationally identify novel miRNAs. Our algorithm uses very few positive samples, requires no negative sample and does not rely on genome annotation. Secondly, I will present our work on genome-wide characterization of promoters of miRNA genes. It is the first piece of work in this field and has been well accepted by biologists. Moreover, I will discuss module discovery in miRNA regulatory networks. Modularity is one of the most prominent properties of real-world complex networks including biological networks. Here, I will address the issue of module identification in bipartite networks and report a novel algorithm especially suited for module detection in them. I analyzed the modules in the miRNA regulatory networks by formulating them into bipartite networks. Finally, I will conclude with an overview of my research interests and plan of my future directions. Bio: Xuefeng Zhou is a fifth-year graduate student in the Department of Computer Science and Engineering at Washington University in Saint Louis. His advisor is Dr. Weixiong Zhang. He received his MS degree in computer science from Illinois Institute of Technology in 2003. Before that, he obtained his MS degree in biochemistry and molecular biology from Peking Union Medical College, and BS degree in biology from Peking (Beijing) University. He expects to receive his Ph.D degree in computer science in August, 2009. His primary research interests lie in bioinformatics/computational biology and data mining, and his current research focuses on computational studies of different aspects of miRNAs as well as other small RNAs. > 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/20090430/a0fe2421/attachment.htm