Cornell University Kavli Institute at Cornell

• About
• Events
• People
• News
• Home

Invited Speakers and Talk Abstracts

[Download Speaker abstracts and bios]

Steven Johnson, Applied Mathematics, MIT
FFTW and Software Architectures to Mitigate Hardware Complexity

Fast Fourier transforms (FFTs) lie at the heart of many problems in scientific computing and engineering. Despite nearly 40 years of scrutiny, they have proved surprisingly difficult to implement efficiently. FFTs almost never achieve the peak CPU speed except for very small problems, with straightforward implementations worse by an order of magnitude over vendor-optimized programs. Obtaining near-optimal performance without laborious machine-specific tuning is a pervasive problem in computing. We believe that a stable and practical solution may be possible by asking: what is the smallest collection of simple algorithmic fragments whose compositions span the optimal algorithm on as many computer architectures as possible. We describe an implementation of this strategy—combining self-optimization, code generation, and recursive cache-oblivious algorithms—in the context of FFTW, an open Fourier transform (FFT) library (www.fftw.org).

Jon Kleinberg, Computer Science, Cornell University
Information Dynamics and Network Structure

With the rapidly growing complexity of interaction patterns in computing systems, the need for theories of information flow in such systems becomes increasingly pressing. We discuss ways to formulate such theories using probabilistic models of networks, which seek to capture properties of network structure, the dynamics of the processes operating on the network, and the feedback between structure and dynamics. An essential aspect of these models is the balance between "short-range" and "long-range" interactions in the underlying network; small shifts in this balance can affect the spread of information in unexpected ways.

Roman Lysecky, Electrical and Computer Engineering
University of Arizona
Autonomously Adaptive Computing

Over the past several decades, incremental advances in computing technology have been enabled by a languages-compilers-architectures ecosystem supported by the software binary. However, as integrated circuit capacities race towards trillions of transistors on a single chip, scalability, reliability, dynamicity, and productivity concerns threaten to disrupt this ecosystem. A key challenge in moving forward is in extending this ecosystem in the presence of these unavoidable design concerns and constraints. We provide an overview of work in adaptive computing that aims to autonomously optimize the execution of a software application at runtime without requiring any designer effort.

Marc Mézard, CNRS-Université Paris Sud
Message Passing Strategies

Message passing algorithms are particularly efficient for dealing with large-scale graphical models, with applications ranging from constraint satisfaction to data clustering. In recent years we have learned how to get from "simple" belief propagation message passing towards more sophisticated algorithms that are able to handle some correlations among variables. The talk will survey the present status of message passing algorithms, their strength, and try to list some of the obstacles they will have to overcome in the next few years.

Hans Mooij, Delft Technical University
Kavli Institute of Nanoscience
Quantum Computation with Nanocircuits

The transport properties of even the smallest practical electronic devices are based on statistical averages over many electronic quantum states. Recently, strong interest has grown in coherent manipulation of individual quantum states. Superposition and entanglement are the key aspects for this field, known as quantum information processing. Theory shows that a quantum computer with quantum bits and quantum gates can be exponentially faster than any conventional computer, for certain specific mathematical operations. Best results so far have been obtained with trapped ions. Solid-state nanofabrication can in principle be used to realize large soli- state circuits that show the same coherent quantum effects. In recent years, important progress has been made in this direction, in particular with individual electrons in semiconductor quantum dots and with superconducting nanocircuits. The time scale for real applications is very long.

Ravi Nair, T.J. Watson Research Center, IBM
Approximate Computing

There is an unprecedented amount of data being produced in the world today. Yet, cost and energy considerations are limiting a corresponding growth in the compute capability needed to process and analyze this data in a conventional manner. There are significant computational and energy efficiencies to be gained by relaxing the expectation of preciseness in today’s computational models and moving to a more approximate computing model. Worldwide, there is more computing of the sort where the results are ephemeral, and where there is a greater acceptance of approximate results. Such results are acceptable even in enterprise computing, where activities such as decision support, search, and data-mining are consuming increasingly greater cycles compared to traditional activities. This talk examines the implications of approximate computing on the exploitation of new technology and on the design of future systems.

Olaf Sporns, Psychological and Brain Sciences
Indiana University, Bloomington
Brain Networks for Efficient Computation

The relationship between structure and function is of central importance for all biological systems including the brain. My talk will be about emerging links between the connectivity structure of the brain and its functional dynamics. Large-scale anatomical brain networks exhibit a number of topological features, including small-world attributes, modularity and community structure, and hubs. How do these structural attributes relate to functional characteristics of brain networks, to their dynamic patterns, to their processing power, robustness, or capacity to support flexible behavior? I will review recent work on complex brain networks that aims to identify how brain networks are organized and how they process and integrate information. Recently, these efforts have included the design of a comprehensive structural and dynamic model of the human cerebral cortex.

About the Symposium

At the nanoscale, one can fit trillions of devices in a chip-scale area. This is many orders of magnitude more than we can accomplish today. The Symposium on Computing Challenges brings together people with hardware, software, neuroscience, physics, and mathematics perspectives to discuss the key problems and the possible approaches to solutions for effectively harnessing trillions of devices that are made possible by use of the nanoscale. [Symposium website]

Organizing Committee

Sandip Tiwari, Electrical and Computer Engineering, Cornell
Rajit Manohar, Electrical and Computer Engineering, Cornell
José Martínez, Electrical and Computer Engineering, Cornell

Co-sponsor

National Nanotechnology Infrastructure Network (NNIN)