http://research.microsoft.com/apps/dp/vi/videos.aspx Watch the latest lectures from Microsoft Research © 2013 Microsoft Corporation. All rights reserved. en-US Fri, 17 May 2013 10:20:34 GMT Fri, 17 May 2013 10:20:34 GMT [Speaker: Narayan Hegde] Most real-world applications use data structures which grow dynamically at runtime. These programs use the dynamic array (vector) implementation available in most programming languages, for example, STL with C++. These libraries handle dynamic growth of data and allow data to be accessed in O(1) time, using arrays. However, in existing implementations of multidimensional dynamic arrays, rows are allocated non-contiguously on the heap, so that each row could grow dynamically. This array of arrays implementation needs multiple memory accesses to read a single data location. We implement a high performance dynamic array (VARRAY) using contiguous multidimensional arrays, so that each location can be accessed via a single memory read. VARRAY improves TLB and cache reuse and other compiler optimizations such as tiling. VARRAY also results in efficient generation of data communication code for distributed memory. Preliminary results using this new data structure show speedup of 12% for 2d arrays and 17% for 3d arrays for programs in polybench. Using array implementation of dynamic arrays, also exposes parallelization opportunity, whose data layout changes dynamically at runtime. Using array notation, we transform PUSH/POP operations on dynamic arrays as read/write operations on array to detect the dependencies in program. We assume non-affine access and runtime dependent control flow, to model irregular parallelism. We then parallelize the data structure operations At runtime we use 2 techniques to paralleize a sequence iterations on dynamic array Loop Strip-mining and Control flow decoupling to identify the true location of update for each PUSH/POP operation on dynamic array. Speculation of push/pop edit locations. These techniques can remove the inherent dependencies on PUSH-POP operations and improve the program performance. http://research.microsoft.com/apps/video/default.aspx?id=192762 Narayan Hegde Science and Technology Thu, 16 May 2013 04:00:00 GMT [Speaker: Alexander Holroyd] Cellular automata display an extraordinary range of behavior, ranging from very simple to apparently chaotic, with many cases in between. Perhaps the most interesting rules are those that yield multiple behavior types from different initial conditions - this is common even for one-dimensional rules started from finitely-supported seeds. If a rule yields chaos from some initial conditions, it is tempting to conclude by analogy with the second law of thermodynamics that chaos should be prevalent from almost all initial conditions. For a certain natural class of rules, we prove that the opposite holds: typical (i.e. random) initial seeds self-organize into predictable (but non-trivial) evolution, while exceptional seeds generate more complicated behavior, including apparent chaos. The key technique is application of percolation arguments to the highly non-independent setting of space-time configurations of cellular automata. http://research.microsoft.com/apps/video/default.aspx?id=192704 Alexander Holroyd Science and Technology Wed, 15 May 2013 22:30:00 GMT [Speaker: Mario Livio] We all make mistakes, because nobody is perfect—and that includes five of the greatest scientists in history: Darwin, Lord Kelvin, Linus Pauling, Fred Hoyle, and Einstein. But … their mistakes helped advance science—and science thrives on error, advancing when erroneous ideas are disproven. Mistakes in any discipline that is based on creative thinking and innovation are not only inevitable, they are an essential part of progress. Breakthroughs require the willingness to embrace risks and to accept errors as potential portals of discovery. http://research.microsoft.com/apps/video/default.aspx?id=192706 Mario Livio Science and Technology Wed, 15 May 2013 20:30:00 GMT [Speaker: John Krakauer] Society arguably admires motor skill above all other human achievements if the popularity of sports and action movies are anything to go by. Despite this admiration, society also seems to rank the theoretical over the practical. Here I will discuss what recent research, including our own, has revealed about motor learning and motor skill, both at the behavioral and neural level. I will then discuss what we need to do to move our understanding forward. http://research.microsoft.com/apps/video/default.aspx?id=192708 John Krakauer Science and Technology Tue, 14 May 2013 17:30:00 GMT [Speaker: Evgeny Morozov] Our society is at a crossroads. Smart technology is transforming our world, making many aspects of our lives more convenient, efficient and - in some cases – fun. Better and cheaper sensors can now be embedded in almost everything, and technologies can log the products we buy and the way we use them. Technology, author Evgeny Morozov proposes, can be a force for improvement - but only if we abandon the idea that it is necessarily revolutionary and instead genuinely interrogate why and how we are using it. From urging us to drop outdated ideas of the Internet to showing how to design more humane and democratic technological solutions, Morozov presents the reasons why we will always need to consider the consequences of the way we use technology. http://research.microsoft.com/apps/video/default.aspx?id=192625 Evgeny Morozov Science and Technology Mon, 13 May 2013 20:30:00 GMT [Speaker: Winter Mason] Although social conventions are a powerful guide for behavior, the way they emerge in communities is not well understood. We focus on competing conventions for attributing reposts to the original source on Twitter. We analyze a near-complete history of tweets, and observe how the conventions emerged and spread through the network of Twitter users. Initially the most successful conventions were borrowed from natural language ("via" and "retweeting"), but over time a community-specific convention came to dominate ("RT"). Additional evidence suggests this specificity to the community, along with efficiency of communication and the timing of the adoption, are key variables in the acceptance of the convention. http://research.microsoft.com/apps/video/default.aspx?id=192624 Winter Mason Science and Technology Mon, 13 May 2013 20:30:00 GMT [Speaker: David Walker] Modern computer networks perform a bewildering array of tasks, from routing and traffic monitoring, to access control and server load balancing. Moreover, historically, managing these networks has been hideously complicated and error-prone, due to a heterogeneous mix of devices (e.g., routers, switches, firewalls, and middleboxes) and their ad hoc, closed and proprietary configuration interfaces. Software-Defined Networking (SDN) is poised to change this state of affairs by offering a clean, simple and open interface between network devices and the software that controls them. In particular, many commercial switches now support the OpenFlow protocol, and a number of campus, data-center, and backbone networks have deployed the new technology. However, while SDN makes it possible to program the network, it does not make it easy: The first generation of SDN controllers offered application developers the "assembly language" of network programming platforms. To reach SDN's full potential, research in programming languages and compilers is desperately needed. In this talk, I discuss our work to date on the Frenetic project, which involves the design of language, compiler and run-time system support for SDN programming. Our languages allow programmers to work declaratively, specifying global behaviors of their network at a high level of abstraction. The compiler and run-time system take care of the tedious details of implementing these high-level policies using the OpenFlow protocol. A key strength of our design is its support for modular programming. Complex network applications can be decomposed in to logical subcomponents — an access control policy, a load balancer, a traffic monitor, a router — and coded independently. Frenetic's rich combinator library makes it possible to stitch such components back together to form a fully functioning whole. In this talk, we will discuss our latest design ideas, including technology that allows programmers to define abstract, virtual networks and isolated network slices. We will also touch on the semantics of Frenetic, explain some of its key properties, and describe a few of the algorithms needed to implement it. http://frenetic-lang.org http://research.microsoft.com/apps/video/default.aspx?id=192626 David Walker Science and Technology Mon, 13 May 2013 17:30:00 GMT [Speaker: Alexander G. Schwing] Sensors acquire an increasing amount of diverse information posing two challenges. Firstly, how can we efficiently deal with such a big amount of data and secondly, how can we benefit from this diversity? In this talk I will first present an approach to deal with large graphical models. The presented method distributes and parallelizes the computation and memory requirements while preserving convergence and optimality guarantees of existing algorithms. I will demonstrate the effectiveness of the approach on stereo reconstruction from high-resolution imagery. In the second part I will present a unified framework for structured prediction with latent variables which includes hidden conditional random fields and latent structured support vector machines as special cases. This framework allows to linearly combine different sources of information and I will demonstrate its efficacy on the problem of estimating the 3D room layout given a single image. For the latter problem, I will introduce a globally optimal yet efficient inference algorithm based on branch-and-bound. http://research.microsoft.com/apps/video/default.aspx?id=192627 Alexander G. Schwing Science and Technology Thu, 09 May 2013 20:30:00 GMT [Speaker: Peter Organisciak] A growing online phenomenon is that of crowdsourcing, where groups of disparate people, connected through technology, contribute to a common product. It refers to the collaborative possibilities of a communications medium as flexible and as populated as the Internet. If many hands make light work, crowdsourcing websites show how light the work can be, breaking tasks into hundreds of pieces for hundreds of hands. Building from the growing body of research in the area including the author’s work on crowd motivations, this talk outlines the considerations necessary in enriching projects through crowdsourcing and discusses ways to interpret the resulting data http://research.microsoft.com/apps/video/default.aspx?id=192541 Peter Organisciak Science and Technology Thu, 09 May 2013 20:30:00 GMT [Speaker: Milind Kulkarni] Over the past several decades of compiler research, there have been great successes in automatically enhancing locality for regular programs, which operate over dense matrices and arrays. Tackling locality in irregular programs, which operate over pointer-based data structures such as trees and graphs, has been much harder, and has mostly been left to ad hoc, application specific methods. In this talk, I will describe efforts by my group to automatically improve locality in a particular class of irregular applications, those that traverse trees. The key insight behind our approach is an abstraction of data structure traversals as operations on vectors. This abstraction lets us design transformations, predict their behavior and determine their correctness. I will present two specific transformations we are developing, "point blocking" and "traversal splicing," and show that they can deliver substantial performance improvements when applied to several real-world irregular kernels. http://research.microsoft.com/apps/video/default.aspx?id=192538 Milind Kulkarni Science and Technology Thu, 09 May 2013 17:30:00 GMT [Speaker: Pedro Domingos] Big data makes it possible in principle to learn very rich probabilistic models, but inference in them is prohibitively expensive. Since inference is typically a subroutine of learning, in practice learning such models is very hard. Sum-product networks (SPNs) are a new model class that squares this circle by providing maximum flexibility while guaranteeing tractability. In contrast to Bayesian networks and Markov random fields, SPNs can remain tractable even in the absence of conditional independence. SPNs are defined recursively: an SPN is either a univariate distribution, a product of SPNs over disjoint variables, or a weighted sum of SPNs over the same variables. It's easy to show that the partition function, all marginals and all conditional MAP states of an SPN can be computed in time linear in its size. SPNs have most tractable distributions as special cases, including hierarchical mixture models, thin junction trees, and nonrecursive probabilistic context-free grammars. I will present generative and discriminative algorithms for learning SPN weights, and an algorithm for learning SPN structure. SPNs have achieved impressive results in a wide variety of domains, including object recognition, image completion, collaborative filtering, and click prediction. Our algorithms can easily learn SPNs with many layers of latent variables, making them arguably the most powerful type of deep learning to date. (Joint work with Rob Gens and Hoifung Poon.) http://research.microsoft.com/apps/video/default.aspx?id=192562 Pedro Domingos Science and Technology Wed, 08 May 2013 20:00:00 GMT [Speaker: Christophe Dubach] Computing systems have become increasingly complex and difficult to program, in particular with the emergence of heterogeneous hardware. For instance it is now common to see GPUs (Graphic Processing Units) used for general purpose computation in data centres or supercomputers. As a result, achieving high performance for such complex systems is an extremely challenging task. This problem is further exacerbated with each new generation of hardware, which means that software written and tuned for today’s systems needs to be adapted frequently to keep pace with ever changing hardware. In this talk I will go through examples of my prior work that attempt to address some of these issues. First I will be discussing the use of machine-learning and automated techniques to build portable optimising compilers. Then I will talk about high-level code generation and optimisations for multicore CPUs and GPUs. Finally, I will describe some of the recent research directions I am exploring related to performance portability for heterogeneous systems. http://research.microsoft.com/apps/video/default.aspx?id=192478 Christophe Dubach Science and Technology Tue, 07 May 2013 10:00:00 GMT [Speaker: Ivana Tošic] Recent development of 3D technologies and depth sensing devices have posed new challenges in processing of depth maps, which are crucial elements in 3D rendering and scene analysis. Typical image processing approaches to these challenges rely on transformations to appropriate representations (e.g., wavelets). However, because of the differences between image and depth statistics, existing image representations might not generalize well to efficiently represent structures in depth maps. One way to develop representations of data with unknown statistics is to learn them from a large database of examples. I will first present a new method for learning dictionaries of waveforms in which depth maps have sparse linear decompositions. The proposed method differs from existing approaches because it is robust to spatially varying noise typical for depth measurements. The effectiveness of this method will be demonstrated on denoising of maps obtained from depth sensors. I will then introduce an algorithm for learning dictionaries that encompass two modalities of 3D scenes: image intensity and depth information. When trained on data from hybrid image-depth sensors, these representations converge to a set of related features, such as pairs of depth and intensity edges or image textures and depth slants. I will conclude by showing some depth inpainting examples based on these hybrid representations. This is joint work with Bruno A. Olshausen, Benjamin J. Culpepper and Sarah Drewes. http://research.microsoft.com/apps/video/default.aspx?id=192425 Ivana Tošic Science and Technology Mon, 06 May 2013 20:30:00 GMT [Speaker: Margaret Mitchell] In this talk, I will be discussing the connection between vision and language, particularly focusing on how to generate human-like reference to visible objects. Some of the findings in this talk have been used to automatically generate descriptions of images (EACL 2012), order descriptive modifiers before a noun (ACL 11), and approximate human preferences for describing color and size (ENLG 2011, NAACL 2013). Evaluating automatically generated text offers interesting challenges, particularly when aiming to capture speaker variation, and I will spend some time discussing how best to measure the naturalness of generated descriptions against a corpus of human-produced descriptions. http://research.microsoft.com/apps/video/default.aspx?id=192427 Margaret Mitchell Science and Technology Mon, 06 May 2013 17:30:00 GMT [Speaker: Gabriel Dos Reis] Software has become critical to nearly every aspect of our civilization. Consequently, the complexity of our tools and our needs for dependability have increased immensely. Programmers need scalable tools and methodogies to keep complexity in check. Generic programming, with roots in computer algebra and symbolic mathematics, is one of the promising approaches to scalable and dependable software contruction. This talk explores recent accomplishments and lessons learned from fruitful interactions between tools support for principled programming and computational mathematics. Examples come from modern C++, OpenAxiom, symbolic mathematics, algorithmic differentiation and beyond. I will touch upon on-going projects, challenges ahead, and possible paths to solutions. http://research.microsoft.com/apps/video/default.aspx?id=192429 Gabriel Dos Reis Science and Technology Thu, 02 May 2013 17:30:00 GMT [Speaker: Jens Teubner] The classical wisdom is that hashing is preferred method to implement joins in main memory. But this wisdom is now many years, if not decades, old and hardware has evolved considerably in the meantime. In this talk I will discuss join strategies for execution in main memory, including hash and sort-merge variants. The runtime characteristics of either strategy depends critically on a suitable implementation that respects the intricacies of modern hardware architectures. I will show how hash and sort-merge joins can be implemented to maximally benefit from hardware features like vector processing (SIMD), multi-level caches, multi-core parallelism, or NUMA-style memory arrangement. And I will point to pitfalls that could mis-guide conclusions about the “best” join implementation strategy. The join implementations discussed outperform the state of the art in join processing by several factors. Experiments on modern hardware platforms indicate that the sort-merge strategy is about to surpass hashing in upcoming hardware architectures—which might put classical wisdom upside-down. http://research.microsoft.com/apps/video/default.aspx?id=192431 Jens Teubner Science and Technology Wed, 01 May 2013 20:30:00 GMT [Speaker: Dave Wecker] Compilers and computer-aided design tools will be essential for quantum computing. At this event, Dave Wecker will present a computer-aided design flow, called LIQUi|⟩, which transforms a high-level language program, representing a quantum algorithm, into a technology-specific implementation. He’ll trace the significant steps in this flow and illustrate the transformations to the representation of the quantum program. Dave will also describe, in detail, the low-level quantum language and the quantum simulator within this design flow. LIQUi|⟩ is an attempt to remedy many of the difficulties in quantum programming and simulation (namely difficulties in debugging, visualization, and exponential growth in memory), as well as make it possible for a wider audience to design, implement, and test quantum algorithms. It provides a functional language (based on F#) that is efficient for programming and allows a combination of classical and quantum operations. In addition, this advanced memory model allows simulation of a large numbers of qubits, and also uses parallel hardware for processing efficiency. Dave will demonstrate the capabilities of LIQUi|⋅⟩ and simulate several quantum algorithms. Shor’s algorithm has been fully implemented in LIQUi|⋅⟩; we can factor numbers from 15 to 8189 on a standard desktop and display the corresponding circuit diagrams. http://research.microsoft.com/apps/video/default.aspx?id=192868 Dave Wecker Science and Technology Tue, 30 Apr 2013 09:00:00 GMT [Speakers: Eric Horvitz, Iain Buchan, Lionel Tarassenko, and Michel Cosnard] Jeannette Wing, Microsoft Research chairs a unique opportunity during the summit where a panel of experts take a step back from the details of the current research problems and look into the future not only in the topic of “(Big) Data” and what it opportunities it offers to machine learning, but more broadly. Panelists: Eric Horvitz, Microsoft Research, Michel Cosnard, INRIA, Iian Buchan, University of Manchester, Lionel Tarassenko, University of Oxford http://research.microsoft.com/apps/video/default.aspx?id=192057 Eric Horvitz; Iain Buchan; Lionel Tarassenko; Michel Cosnard Science and Technology Wed, 24 Apr 2013 22:00:00 GMT [Speaker: Matthew Green] Bitcoin is the first e-cash system to see widespread adoption. While Bitcoin offers the potential for new types of financial interaction, it has significant limitations regarding privacy. Specifically, because the Bitcoin transaction log is completely public, users’ privacy is protected only through the use of pseudonyms. In this talk we discuss Zerocoin, a cryptographic extension to Bitcoin that augments the protocol to allow for fully anonymous currency transactions. Our system uses standard cryptographic assumptions and does not introduce new trusted parties or otherwise change the security model of Bitcoin. We detail Zerocoin’s cryptographic construction, its integration into Bitcoin, and examine its performance both in terms of computation and impact on the Bitcoin protocol. http://research.microsoft.com/apps/video/default.aspx?id=192058 Matthew Green Science and Technology Wed, 24 Apr 2013 20:30:00 GMT [Speaker: Yair Weiss] Compressed sensing (CS) refers to a branch of applied mathematics which is based on the surprising result whereby signals that are exactly “k-sparse” (i.e. can be represented by at most k nonzero coefficients in some basis) can be exactly reconstructed using a small number of random measurements. Since natural images tend to be sparse in the wavelet basis, one of the motivating examples of CS has always been to reconstruct high resolution images from a small number of random measurements. Unfortunately, there are some significant deviations between the way that natural images behave and the assumptions of the dramatic theorems, and in fact random projections perform quite poorly when applied to real images. I will describe an alternative theory, which we call “Informative Sensing”, that seeks a small number of projections that are maximally informative given a known distribution over signals. I will show experimental results demonstrating that the informative projections indeed outperform random projections, but that the savings relative to more standard imaging methods are altogether rather modest. Joint work with Hyun Sung Chang and Bill Freeman. http://research.microsoft.com/apps/video/default.aspx?id=192246 Yair Weiss Science and Technology Wed, 24 Apr 2013 20:00:00 GMT