General Purpose, Low Power Supercomputing Using Reconfiguration
Google engEDU
56 min – Feb 28, 2006

Google TechTalks
February 28, 2006

Prof. Bob Brodersen

ABSTRACT:
The ability of FPGA technology to exploit the advances in IC fabrication technology has resulted in the present situation in which a FPGA computing fabric is the most power and area efficient approach for general purpose parallel computing. This has occurred because the Von-Neumann processor architectures are now power limited and can no longer fully exploit the technology advances (thus the move to multi-cores). Hardware composed of arrays of FPGA’s and memory has been design that achieves a TeraOp/second of performance per board with over an order of magnitude higher efficiency for the computation per unit power over conventional microprocessors. To achieve these results, however, requires a high level of parallelism in the application program, which is typically not exposed in sequential programming languages. Even worse for application programmers, has been the low level of abstraction of FPGA hardware, which requires the user to be a hardware expert. It is believed that for any application that can be parallelized and streamed will presently achieve orders of magnitude speed-up for the same power and cost and even more importantly will have a power efficiency which will improve exponentially in each subsequent IC technology node. Read the rest of this entry »

Decision Making and Chance
Google engEDU
59 min – Sep 17, 2006

Google Tech Talks
September 17, 2006

Dr. Mike Orkin is a Managing Scientist at Exponent, a publicly traded scientific consulting company headquartered in Menlo Park. Mike has numerous research publications in game theory and probability theory and has written data mining and simulation software. He is a nationally known authority on odds and gambling games and has appeared on numerous TV and radio shows to discuss gambling and odds, including CNN, NBC’s Dateline and ABC’s World News Tonight.

ABSTRACT
Certain gambling games, such as roulette and craps, are games of pure chance: In repeated play, luck disappears, and the persistent gambler will go broke. Other gambling activities, such as betting on sports or the stock market, may involve an element of skill. One way to measure this is to compare the results of a gambling strategy with chance: A skillful strategy should produce long-run results that are better than would be achieved by someone who is just guessing. One can also compare a gambler’s losses with chance to see if the gambler is doing worse than chance would allow. I will discuss two recent projects that illustrate these concepts:

• Automated data mining software discovers that the Baltimore Ravens are 17-3 versus the point spread when they lost their previous game and their opponents played their previous game on the road. Do situations like this give clever gamblers an edge or are such strong win-loss records merely random flukes?

• A gambler loses $30 million betting at an online casino. Is it possible to lose this much just by chance or is the gambler being cheated? Or maybe the gambler is part of a money laundering scheme. Read the rest of this entry »

Incremental Bayesian Networks for Natural Language Parsing
Google engEDU
1 hr 1 min – Aug 13, 2007

Google Tech Talks
August 13, 2007

ABSTRACT

Natural language parsing is a particularly challenging structure prediction problem, due to the large space of output structures and the complex nature of the statistical dependencies between features of the output structures. Typically these statistical dependencies are specified by hand, but recently there has been interest in using latent variables to induce them automatically. In this talk I will present a framework for structure prediction with latent variables based on a form of Dynamic Bayesian Network called Incremental Sigmoid Belief Networks (ISBNs), and illustrate how it can be applied to parsing. Approximations to ISBNs have achieved state-of-the-art results on Penn Treebank parsing and on dependency parsing for a variety of languages.

ISBNs are designed for non-Markovian problems such an parsing, where the structure of the statistical dependencies is a function of the output structure. Exact inference in ISBNs cannot be done efficiently (as with many complex graphical models), but they are designed to allow efficient approximations. Such approximations include my previous neural network architecture for statistical parsing, and an incremental mean field approximation. The mean field approximation demonstrates that a more accurate approximation does lead to a more accurate parser, but the neural network approximation is much faster and achieves close to the same accuracy. Read the rest of this entry »

NASA’s "Beyond Einstein" Program: Exploration at the Limits of Space & Time

53 min – Aug 30, 2007

Google Tech Talks
August, 30 2007

ABSTRACT

Albert Einstein’s General Theory of Relativity predicted results that were so incredible that even he did not accept them: space is expanding from a Big Bang, space itself contains an energy that is pulling the Universe apart from within, and deep chasms of gravity called black holes actually exist. Astonishingly, all of these wild ideas are now known to be true. But now we need to build on Einstein’s work to take the next step — to study the underlying physics of the very phenomena that came out of his theories. NASA’s Beyond Einstein program consists of a series of space missions, large and small, that push Einstein’s theories to their limits, using increasingly more sensitive probes. The two flagship missions now in development, Constellation-X and LISA, will explore extremes of space, measuring X-rays and gravitational waves. The smaller missions, the Einstein probes, will target specific science questions such as "What is Dark Energy?" and "What powered the Big Bang?"

Speaker: Randall Smith
Randall Smith is an X-ray Astrophysicist at the Johns Hopkins University and NASA/Goddard Space Flight Center. He worked at the Chandra X-ray Center for six years after the launch of the Chandra X-ray Observatory before moving to JHU & NASA to work with the Suzaku X-ray Telescope. He is a member of the Cons… Read the rest of this entry »

Scalability and Efficiency on Data Mining Applied to Internet Applications
Google engEDU
43 min – Aug 16, 2007

Google Tech Talks
August 16, 2007

ABSTRACT

The Internet went well beyond a technology artefact, increasingly becoming a social interaction tool. These interactions are usually complex and hard to analyze automatically, demanding the research and development of novel data mining techniques that handle the individual characteristics of each application scenario. Notice that these data mining techniques, similarly to other machine learning techniques, are intensive in terms of both computation and I/O, motivating the development of new paradigms, programming environments, and parallel algorithms that support scalable and efficient applications. In this talk we present some results that justify not only the need for developing these new techniques, as well as their parallelization.

Wagner Meira Jr. obtained his PhD from the University of Rochester in 1997 and is currently Associate Professor at the Computer Science Department at Universidade Federal de Minas Gerais, Brazil. His research focuses on scalability and efficiency of large scale parallel and distributed systems, from massively parallel to Internet-based platforms, and on data mining algorithms, their parallelization, and application to areas such as information retrieval, bioinformatics, and e-governance. Read the rest of this entry »