- talks - lists.cs.princeton.edu

Colloquium Speaker: Andrej Karpathy Monday April 11, 12:30pm
by Nicole E. Wagenblast 07 Apr '16

07 Apr '16

Colloquium Speaker Andrej Karpathy, Stanford University Monday, April 11- 12:30pm Computer Science 105 Connecting Images and Natural Language Intelligent agents require the ability to perceive their environments, understand their high-level semantics, and communicate with humans. While computer vision has recently made great strides on visual recognition tasks, the predominant paradigm is to predict one or more fixed visual categories for each image. I will describe a line of work that significantly expands the vocabulary of our computer vision systems and allows them to express visual concepts in natural language, such as “a picture of a girl playing with a stack of legos”, or “a couple holding hands and walking on a beach”. In particular, the final model can take an image and both detect and describe in natural language all of its salient regions. My modeling techniques draw on recent advances in Deep Learning that allow us to construct and train neural networks with hundreds of millions of neurons that take raw images and map them directly to natural language sentences. I will show that the model generates qualitatively compelling results and quantitative evaluation and control experiments demonstrate the strength of this approach with respect to simpler baselines and previous methods. Andrej Karpathy is a Computer Science Ph.D. candidate at Stanford University working with Prof. Fei-Fei Li. He received his M.S. from University of British Columbia in Computer Science and his B.S from University of Toronto in Computer Science and Physics. He is interested in the intersection of computer vision, natural language processing and reinforcement learning with the aim of developing agents who can intelligently interact with humans in dynamic environments. His work was featured in the New York Times and MIT Technology Review. He helped develop and instruct a new Computer Science class at Stanford on Convolutional Neural Networks for Visual Recognition. In his spare time he develops Deep Learning libraries in Javascript and maintains websites that support more efficient meta-research.

1 0

Colloquium Speaker Roger Grosse, TODAY April 7th- 12:30pm
by Mitra D. Kelly 07 Apr '16

07 Apr '16

Colloquium Speaker Roger Grosse, the University of Toronto April 7th- 12:30pm Computer Science 105 Exploiting compositionality to explore a large space of model structures I will present flexible algorithms for model discovery and model fitting which apply to broad, open-ended classes of models, yet which also incorporate model-specific algorithmic insights. First, I will introduce a framework for building probabilistic models compositionally out of common modeling motifs, such as clustering, sparsity, and dimensionality reduction. This compositional framework yields a variety of existing models as special cases. We can flexibly perform posterior inference across this large, open-ended space of models by composing sophisticated inference algorithms carefully designed for the individual modeling motifs. An automatic structure search procedure over this space of models yields sensible analyses of datasets as diverse as motion capture, natural image patches, and Senate voting records, all using a single software package with no hand-tuned metaparameters. Applying a similar compositional structure search procedure to Gaussian Process models yields interpretable decompositions of diverse time series datasets and enables automatic generation of natural language reports. Finally, compositional structure search depends crucially on the estimation of intractable likelihoods. I will briefly outline an approach for obtaining precise likelihood estimates with rigorous tail bounds by sandwiching the true value between stochastic upper and lower bounds. BIO: Roger Grosse is a Postdoctoral Fellow in the University of Toronto machine learning group. He received his Ph.D. in computer science from MIT under the supervision of of Bill Freeman. He is a recipient of the NDSEG Graduate Fellowship, the Banting Postdoctoral Fellowship, and outstanding paper awards at the International Conference of Machine Learning (ICML) and the Conference for Uncertainty in AI (UAI). He is also a co-creator of Metacademy, an open-source web site for developing personalized learning plans in machine learning and related fields. Mitra Kelly Academic Secretary Princeton University Computer Science Dept 35 Olden Street Princeton NJ 08540 mkelly(a)cs.princeton.edu <mailto:mkelly@cs.princeton.edu> 609-258-4562

1 0

Colloquium Speaker Yin-Tat Lee, TODAY April 5th- 12:30pm
by Mitra D. Kelly 05 Apr '16

05 Apr '16

Colloquium Speaker Yin-Tat Lee, Massachusetts Institute of Technology April 5th, 12:30pm Computer Science 105 Title: Faster algorithms for fundamental convex problems and their applications in combinatorial optimization Abstract: Convex optimization has been studied extensively and is a prominent tool in various areas such as combinatorial optimization, data analysis, operations research, and scientific computing. Each field has developed specialized tools including data structures, sampling methods, and dimension reduction. In the past several years, I have been combining and improving the optimization techniques from different fields to design faster optimization algorithms. In this talk, I will discuss my work in this direction and illustrate it through my results on linear programming and general convex optimization. In particular, I will present a new algorithm for solving linear programs, which gives the first improvement to the running time for linear programming in 25 years. Then, I will present the first nearly cubic time algorithm for solving general convex optimization problems. Furthermore, I will discuss how these two results can be used to improve the running time of many classical combinatorial problems such as maximum flow and submodular function minimization. This talk will assume no prior knowledge of optimization. Bio: Yin Tat Lee is a Ph.D. candidate in the department of mathematics at the Massachusetts Institute of Technology. He is interested in designing faster algorithms, particularly for problems in optimization. Since he began his Ph.D. in 2012, he has combined ideas from continuous and discrete mathematics to substantially advance the state-of-the-art for solving many fundamental problems in computer science, such as linear programming, maximum flow, and submodular function minimization. He has received a variety of awards, including the Best Student Paper Award at FOCS 2015, Best Paper Award at SODA 2014, Best Paper Award and Best Student Paper Award at FOCS 2014, and Notable Article in Computing in 2014 by Computing Reviews. Mitra Kelly Academic Secretary Princeton University Computer Science Dept 35 Olden Street Princeton NJ 08540 mkelly(a)cs.princeton.edu <mailto:mkelly@cs.princeton.edu> 609-258-4562

1 0

Colloquium Speaker Yin-Tat Lee, Tuesday April 5th- 12:30pm
by Mitra D. Kelly 04 Apr '16

04 Apr '16

Colloquium Speaker Yin-Tat Lee, Massachusetts Institute of Technology April 5th, 12:30pm Computer Science 105 Title: Faster algorithms for fundamental convex problems and their applications in combinatorial optimization Abstract: Convex optimization has been studied extensively and is a prominent tool in various areas such as combinatorial optimization, data analysis, operations research, and scientific computing. Each field has developed specialized tools including data structures, sampling methods, and dimension reduction. In the past several years, I have been combining and improving the optimization techniques from different fields to design faster optimization algorithms. In this talk, I will discuss my work in this direction and illustrate it through my results on linear programming and general convex optimization. In particular, I will present a new algorithm for solving linear programs, which gives the first improvement to the running time for linear programming in 25 years. Then, I will present the first nearly cubic time algorithm for solving general convex optimization problems. Furthermore, I will discuss how these two results can be used to improve the running time of many classical combinatorial problems such as maximum flow and submodular function minimization. This talk will assume no prior knowledge of optimization. Bio: Yin Tat Lee is a Ph.D. candidate in the department of mathematics at the Massachusetts Institute of Technology. He is interested in designing faster algorithms, particularly for problems in optimization. Since he began his Ph.D. in 2012, he has combined ideas from continuous and discrete mathematics to substantially advance the state-of-the-art for solving many fundamental problems in computer science, such as linear programming, maximum flow, and submodular function minimization. He has received a variety of awards, including the Best Student Paper Award at FOCS 2015, Best Paper Award at SODA 2014, Best Paper Award and Best Student Paper Award at FOCS 2014, and Notable Article in Computing in 2014 by Computing Reviews. Mitra Kelly Academic Secretary Princeton University Computer Science Dept 35 Olden Street Princeton NJ 08540 mkelly(a)cs.princeton.edu <mailto:mkelly@cs.princeton.edu> 609-258-4562

1 0

Colloquium Speaker Philipp Kräehenbüehl, today, 12:30pm
by Nicole E. Wagenblast 04 Apr '16

04 Apr '16

Colloquium Speaker Philipp Kräehenbüehl, University of California, Berkeley Monday, April 4, 2016 - 12:30pm Computer Science 105 The many ways to understand the pixels The field of computer vision is arguably seeing one of its most transformative changes in recent history. Convolutional neural networks (CNNs) have revolutionized the field, reaching super-human performance on some long-standing computer vision tasks, such as image classification. The success of these networks is fueled by massive amounts of human-labeled data. However this paradigm does not scale to a deeper and more detailed understanding of images, as it is simply too hard to collect enough human-labeled data. The issue is not that we humans don't understand the image, but we often struggle to convey enough information to successfully supervise a vision system. In this talk I show how computer vision can go beyond massive human supervision. This involves designing better models that deal with fewer labels, exploiting easier and more intuitive annotations, or coming up with novel optimizations to train deep architectures with far fewer human annotations, or even without any at all. I'll focus on three long standing computer vision problems: semantic segmentation, intrinsic image decomposition and dense semantic correspondences. Philipp Krähenbühl is a postdoctoral researcher at UC Berkeley. He received a B.S. in Computer Science from ETH Zurich in 2009, and a PhD in Computer Science from Stanford University in 2014. Philipp's research interests lie in Computer vision, Machine learning and Computer Graphics. He is particularly interested in deep learning, efficient optimization techniques, and structured output prediction.

1 0

Colloquium Speaker Philipp Kräehenbüehl, Monday April 4, 12:30pm- reminder
by Nicole E. Wagenblast 03 Apr '16

03 Apr '16

Colloquium Speaker Philipp Kräehenbüehl, University of California, Berkeley Monday, April 4, 2016 - 12:30pm Computer Science 105 The many ways to understand the pixels The field of computer vision is arguably seeing one of its most transformative changes in recent history. Convolutional neural networks (CNNs) have revolutionized the field, reaching super-human performance on some long-standing computer vision tasks, such as image classification. The success of these networks is fueled by massive amounts of human-labeled data. However this paradigm does not scale to a deeper and more detailed understanding of images, as it is simply too hard to collect enough human-labeled data. The issue is not that we humans don't understand the image, but we often struggle to convey enough information to successfully supervise a vision system. In this talk I show how computer vision can go beyond massive human supervision. This involves designing better models that deal with fewer labels, exploiting easier and more intuitive annotations, or coming up with novel optimizations to train deep architectures with far fewer human annotations, or even without any at all. I'll focus on three long standing computer vision problems: semantic segmentation, intrinsic image decomposition and dense semantic correspondences. Philipp Krähenbühl is a postdoctoral researcher at UC Berkeley. He received a B.S. in Computer Science from ETH Zurich in 2009, and a PhD in Computer Science from Stanford University in 2014. Philipp's research interests lie in Computer vision, Machine learning and Computer Graphics. He is particularly interested in deep learning, efficient optimization techniques, and structured output prediction.

1 0

Colloquium Speaker Yin-Tat Lee, Tuesday April 5th- 12:30pm
by Mitra D. Kelly 31 Mar '16

31 Mar '16

Colloquium Speaker Yin-Tat Lee, Massachusetts Institute of Technology April 5th, 12:30 Computer Science 105 Title: Faster algorithms for fundamental convex problems and their applications in combinatorial optimization Abstract: Convex optimization has been studied extensively and is a prominent tool in various areas such as combinatorial optimization, data analysis, operations research, and scientific computing. Each field has developed specialized tools including data structures, sampling methods, and dimension reduction. In the past several years, I have been combining and improving the optimization techniques from different fields to design faster optimization algorithms. In this talk, I will discuss my work in this direction and illustrate it through my results on linear programming and general convex optimization. In particular, I will present a new algorithm for solving linear programs, which gives the first improvement to the running time for linear programming in 25 years. Then, I will present the first nearly cubic time algorithm for solving general convex optimization problems. Furthermore, I will discuss how these two results can be used to improve the running time of many classical combinatorial problems such as maximum flow and submodular function minimization. This talk will assume no prior knowledge of optimization. Bio: Yin Tat Lee is a Ph.D. candidate in the department of mathematics at the Massachusetts Institute of Technology. He is interested in designing faster algorithms, particularly for problems in optimization. Since he began his Ph.D. in 2012, he has combined ideas from continuous and discrete mathematics to substantially advance the state-of-the-art for solving many fundamental problems in computer science, such as linear programming, maximum flow, and submodular function minimization. He has received a variety of awards, including the Best Student Paper Award at FOCS 2015, Best Paper Award at SODA 2014, Best Paper Award and Best Student Paper Award at FOCS 2014, and Notable Article in Computing in 2014 by Computing Reviews. Mitra Kelly Academic Secretary Princeton University Computer Science Dept 35 Olden Street Princeton NJ 08540 mkelly(a)cs.princeton.edu <mailto:mkelly@cs.princeton.edu> 609-258-4562

1 0

Colloquium Speaker Philipp Kräehenbüehl, Monday April 4, 12:30pm
by Nicole E. Wagenblast 31 Mar '16

31 Mar '16

Colloquium Speaker Philipp Kräehenbüehl, University of California, Berkeley Monday, April 4, 2016 - 12:30pm Computer Science 105 The many ways to understand the pixels The field of computer vision is arguably seeing one of its most transformative changes in recent history. Convolutional neural networks (CNNs) have revolutionized the field, reaching super-human performance on some long-standing computer vision tasks, such as image classification. The success of these networks is fueled by massive amounts of human-labeled data. However this paradigm does not scale to a deeper and more detailed understanding of images, as it is simply too hard to collect enough human-labeled data. The issue is not that we humans don't understand the image, but we often struggle to convey enough information to successfully supervise a vision system. In this talk I show how computer vision can go beyond massive human supervision. This involves designing better models that deal with fewer labels, exploiting easier and more intuitive annotations, or coming up with novel optimizations to train deep architectures with far fewer human annotations, or even without any at all. I'll focus on three long standing computer vision problems: semantic segmentation, intrinsic image decomposition and dense semantic correspondences. Philipp Krähenbühl is a postdoctoral researcher at UC Berkeley. He received a B.S. in Computer Science from ETH Zurich in 2009, and a PhD in Computer Science from Stanford University in 2014. Philipp's research interests lie in Computer vision, Machine learning and Computer Graphics. He is particularly interested in deep learning, efficient optimization techniques, and structured output prediction.

1 0

Subject: Colloquium Speaker Roger Grosse, Thursday April 7th- 12:30pm
by Mitra D. Kelly 31 Mar '16

31 Mar '16

Colloquium Speaker Roger Grosse, the University of Toronto April 7th- 12:30pm Computer Science 105 Exploiting compositionality to explore a large space of model structures I will present flexible algorithms for model discovery and model fitting which apply to broad, open-ended classes of models, yet which also incorporate model-specific algorithmic insights. First, I will introduce a framework for building probabilistic models compositionally out of common modeling motifs, such as clustering, sparsity, and dimensionality reduction. This compositional framework yields a variety of existing models as special cases. We can flexibly perform posterior inference across this large, open-ended space of models by composing sophisticated inference algorithms carefully designed for the individual modeling motifs. An automatic structure search procedure over this space of models yields sensible analyses of datasets as diverse as motion capture, natural image patches, and Senate voting records, all using a single software package with no hand-tuned metaparameters. Applying a similar compositional structure search procedure to Gaussian Process models yields interpretable decompositions of diverse time series datasets and enables automatic generation of natural language reports. Finally, compositional structure search depends crucially on the estimation of intractable likelihoods. I will briefly outline an approach for obtaining precise likelihood estimates with rigorous tail bounds by sandwiching the true value between stochastic upper and lower bounds. BIO: Roger Grosse is a Postdoctoral Fellow in the University of Toronto machine learning group. He received his Ph.D. in computer science from MIT under the supervision of of Bill Freeman. He is a recipient of the NDSEG Graduate Fellowship, the Banting Postdoctoral Fellowship, and outstanding paper awards at the International Conference of Machine Learning (ICML) and the Conference for Uncertainty in AI (UAI). He is also a co-creator of Metacademy, an open-source web site for developing personalized learning plans in machine learning and related fields. Mitra Kelly Academic Secretary Princeton University Computer Science Dept 35 Olden Street Princeton NJ 08540 mkelly(a)cs.princeton.edu <mailto:mkelly@cs.princeton.edu> 609-258-4562

1 0

Colloquium Speaker, Matt Weinberg Today, 12:30pm
by Nicole E. Wagenblast 31 Mar '16

31 Mar '16

Colloquium Speaker Matt Weinberg, Princeton University Thursday, March 31, 2016 - 12:30pm Computer Science 105 Algorithms for Strategic Agents When real people interact with algorithms (e.g. in auctions, crowdsourcing, Bitcoin, etc.), they impose additional objectives beyond simply that the algorithm is correct, fast, and uses little storage. People strategize during these interactions, so algorithms deployed in these settings must be robust against potential manipulation. Additionally, people prefer transparent interactions, so these algorithms must also be as simple as possible. My research addresses these, and other novel challenges that arise when algorithms interact with strategic agents. In this talk, I will focus on one aspect of this agenda and present a new algorithmic framework to solve any optimization problem in a way that is robust to strategic manipulation. I will further apply this framework to extend Myerson's celebrated characterization of optimal single-item auctions to multiple items (Myerson 1981), design mechanisms for job scheduling (Nisan and Ronen 1999), and resolve other problems at the interface of algorithms and game theory. Finally, I will briefly show how strategic considerations motivate nice questions in "traditional" areas of algorithm design as well, and present some of my work in online algorithms, convex optimization, and parallel algorithms. Matt received his PhD in EECS from MIT in 2014, where he was advised by Costis Daskalakis. He is now a postdoc at Princeton University in the Computer Science department. His research focuses on designing algorithms that address constraints imposed by the strategic nature of people that interact with them. For his thesis work on this topic, he received MIT's George M. Sprowls award and the SIGecom Doctoral Dissertation award. Matt received his B.A. in Math from Cornell University.

1 0