This section was written by Associate Editor Jean Thilmany.
SUPERCOMPUTER UPGRADE
Cystorm, Iowa State University’s second supercomputer, has a peak processing rate about five times that of the university’s other super system, CyBlue, but giving us an idea of how fast computer power is advancing, Cystorm didn’t make the cut for the current list of the top 500 computers in the world. CyBlue, on the other hand, was rated 99th in the world when it first booted up three years ago.
The new supercomputer, a Sun Microsystems machine, is capable of a peak performance of 28.16 trillion calculations per second. The supercomputer, which went online in August, will help Iowa State researchers advance their work in materials science, power systems, and systems biology.
“Cystorm is going to be very good for data-intensive research projects,” said Srinivas Aluru, the Ross Martin Mehl and Marylyne Munas Mehl Professor of Computer Engineering and the leader of the Cystorm project. “The capabilities of Cystorm will help Iowa State researchers do new research in their fields.”
CyBlue, an IBM Blue Gene/L supercomputer on campus since early 2006, uses 2,048 processors to do 5.7 trillion calculations per second. Cystorm has 3,200 processor cores, Aluru said.
The new machine also scores high on a test of actual running performance, using the same test that is used to rank the Top 500. It clocked in at 15.44 trillion calculations per second, compared to CyBlue’s 4.7 trillion per second. That measure makes Cystorm 3.3 times more powerful than CyBlue, Aluru said.
Those performance numbers, however, do not earn Cystorm a spot on the current Top 500 list of the world’s fastest supercomputers. The list, compiled by prominent researchers and issued twice a year, confers bragging rights on research institutions and manufacturers, and serves as a tool to track trends in supercomputer performance and architectures. It is published online at www.top500.org.
No. 500 on the current list is an IBM BladeCenter HS21 cluster, operated by Financial Services in the U.K., with a peak performance of 37.64 trillion calculations per second. Its actual running performance, or Rmax, which determines the computer’s world ranking, is 17.09 trillion calculations per second. No. 1 is the U.S. DOE’s Roadrunner, an IBM BladeCenter QS22 cluster with an Rmax of 1,105 teraflops.
IT GETS ME
RESEARCHERS AT OREGON STATE UNIVERSITY in Covallis are working to develop computers that will want to communicate with, learn from, and get to know you.
Such a computer doesn’t just learn from its own experiences; it listens to its users, and then combines what it’s learned with its reasoning algorithms. Then, it changes its programming as a result, said Margaret Burnett, an associate professor of computer science at OSU.
Also, when ordinary users spot errors in the computer’s logic, they should be able to step in and explain directly to the machine the logic it should be using, Burnett said.
Biologists, neuroscientists, engineers, and computer scientists are working toward robots that will learn much as the human brain does and act much as a human does. And they’re calling upon machine-learning algorithms and specialized software and hardware to do so. This month’s Computing takes a look at how some programs and researchers are driving forward machine-learning and cognitive computing.
|
“There are limits to what the computer can do just by its own observations and efforts to learn from experiences,” she said. “It needs to understand not just what it did right or wrong, but why. And for that, it has to continue interacting with human beings and make constant changes in its own programming, based on their feedback.”
Many advanced learning systems begin to customize themselves to the end user from the moment they are delivered. Systems like this are the brains behind spam filters and product recommendations such as: “If you liked this book, you’ll also like this one.”
A lot of these systems use word statistics, set rules, similarities, and other such approaches. But even the most advanced systems only allow a user to tell the computer something is right or wrong. The user is never asked to explain what the real problem is, said Weng-Keen Wong, an OSU assistant professor of computer science.
“We want to develop algorithms that will allow the end user to ask the computer why it did something, read the computer’s response, and then to explain to the computer why that was a mistake,” Wong said. “Ideally, the computer will consider the response and change its programming to perform better in the future. It’s like debugging a program.”
A major part of this challenge, the OSU scientists said, is to create interactive systems that are easy enough to operate that you don’t have to be a computer programmer to run them. That should be possible, they said.
BRAIN IN A BOX
A COMPUTER FUNCTIONS BY ESSENTIALLY consulting a list of programmed algorithms. A human brain does not.
Human and computer are separated by many impossible-to-bridge gulfs, of course, but soon cognition may no longer be among them, according to an unlikely alliance of researchers from fields that include computer science, mathematics, neuroscience, kinematics, and cognitive science. They’re working separately toward computers and robots that reason, learn, and respond much as a human does.
These computers of the not-so-distant future will sense our emotions and maybe even express their own. They’ll adapt to their environments and behave flexibly and intelligently when faced with novel and unexpected situations.
These types of cognitive computing projects have two sides, said Stan Franklin, who describes himself as a mathematician turned computer scientist turning cognitive scientist. He’s a professor in the Cognitive Computing Research Group at the University of Memphis in Tennessee. In order to replicate the brain’s working digitally, scientists need to understand exactly how the brain processes information—a topic long studied and still not defined. Computer scientists look to that cognitive model to build computers and robots with human-like intelligence, perception, and emotion. When the cognitive and the engineering puzzles are solved, and then joined, a computer that thinks, acts, and feels like a person will be possible, Franklin said.
Researchers, including Franklin, are well on their way toward this end, he added.
No one knows exactly how the cognitive computers of the future will look and predictions about how they’ll be used are just that, said Javier Movellan, principal investigator at the Machine Perception Laboratory at the University of California, San Diego. He and his team have spent the last 15 years developing machine-learning interfaces that will power cognitive computers.
With the field still in its infancy, it’s hard to conceive all possibilities for its applications, he said.
His lab is now finding interesting results from its cognitive robots. The computers are essentially babies, and learn just as babies do, he said. They have shown a surprising ability to identify human faces represented in different media.
A computer that can learn about its environment, by taking cues from its surroundings and acting on them, has potentially many more applications than can be wrested from today’s powerful machines, said Martin McGinnity, a professor of intelligent systems engineering at University of Ulster in Northern Ireland. His team’s Sensemaker captures sight, sound, and touch sensory data, and combines them for a whole picture of an environment. The team is made up of biologists, neuroscientists, engineers, and computer scientists.
Such cognitive computers could one day read people’s faces for clues about their emotional state, according to MIT’s Affective Computing Lab, which is also at work on a number of projects that marry emotion and technology. An autistic person could wear a specialized emotion-sensing device to decipher their friends’ and coworkers’ nonverbal messages, for example.
The programs that allow these computers to learn about their environments or register emotion are based on machine-learning algorithms. The algorithms contain rules that allow machines to get better at whatever they’re programmed to learn. Just as we have neurons, the computer programs that act as the brain for Movellan’s baby robots contain little units of information analysis that take in data and then produce a signal that affects the next neuron, or analysis unit, down the line, he said.
While specialized learning algorithms lead to computers that think, albeit primitively, like a human, the programs need to learn about their surroundings the way we do—by piecing together bits of information to get a view of the world. Movellan, for instance, knew that, in order for his robots to interact with humans, they first need to recognize a human face. All other interactions would follow from that initial recognition.
“Rather than figuring out by hand how to solve problems like finding faces, we take many examples of faces and no faces and put them in the computer to learn the difference,” he said.
The researchers connected the algorithm’s growing capabilities with motors and actuators to create robots that behave in an adaptive and useful manner. These first robots were installed in baby dolls.
“One of our experiments with the baby robots was that every once in a while we would touch and play with it, and that was enough for them to learn what humans look like,” Movellan said.
To the team’s surprise the robots quickly learned to pick out human faces so well they could separate the human from the inhuman in comic books and other formats.
The potential applications for this cognitive computing research may still seem a bit cloudy. But, as Movellan said, we can’t yet imagine the full potential.
The cognitive computing field is now just taking off, Franklin said.
WHAT'S THAT?
PEOPLE SEE, HEAR, AND FEEL, and make sense of countless diverse, quickly changing stimuli in our environment seemingly without effort. But doing what our brains do with ease is often an impossible task for computers.
Researchers at the Leipzig Max Planck Institute for Human Cognitive and Brain Sciences and the Wellcome Trust Centre for Neuroimaging in London have now developed a mathematical model that they claim could significantly improve the automatic recognition and processing of spoken language.
In the future, these kinds of algorithms, which imitate brain mechanisms, could help machines perceive the world around them, said Stefan Kiebel from the Leipzig Max Planck Institute for Human Cognitive and Brain Sciences. He’s one of the researchers who helped develop the mathematical model.
Many people have personally experienced how difficult it is for computers to deal with spoken language. Ever phone your telephone company or health insurance provider only to be greeted by an automated telephone system?
You’ll need a great deal of patience when speaking to these types of computers. If you speak just a little too quickly or slowly, if your pronunciation isn’t clear, or if you have music playing in the background, the computerized operator won’t hear you properly.
The reason for their failure to comprehend your words is that until now these computer programs have relied on processes which are particularly sensitive to perturbations, Kiebel said. When computers process language, they primarily attempt to recognize characteristic features in the frequencies of the voice in order to recognize words.
“It’s likely the brain uses a different process,” he said.
The researcher presumes that a different process involves temporal sequences. Music and spoken language, for example, are composed of sequences of different length. According to the scientist’s hypothesis, the brain classifies the various signals from the smallest, fast-changing components—such as single sound units like “e” or “u”—up to big, slow-changing elements—such as the topic being discussed.
“The brain permanently searches for temporal structure in the environment in order to deduce what will happen next,” Kiebel said.
In this way, the brain can often predict the next sound units based on the slow-changing information. Thus, if the topic of conversation is the hot summer, the “su” sound will more likely be the beginning of the word “sun” than of the word “supper.”
To test this hypothesis, the researchers constructed a mathematical model that was designed to imitate, in a highly simplified manner, the neuronal processes that occur during the comprehension of speech. Neuronal processes were described by algorithms that processed speech at several temporal levels. The model succeeded in processing speech; it recognized individual speech sounds and syllables, Kiebel said.
The success, although it is early and with the simplest parts of speech, confirms the hypothesis, Kiebel said. The model provides new approaches for practical applications in the field of artificial speech recognition, he said.
BRIEFLY NOTED
Ansys Inc. of Canonsburg, Pa., has released its HFSS 12.0 software for 3-D full-wave electromagnetic field simulation. The product is part of the developer’s Ansoft suite. /// CAD Schroer Group of Moers, Germany, has released version 4.0 of MPDS4, its plant design and factory layout system. /// Magsoft Corp. of Ballston Spa, N.Y., is now shipping Flux version 10.3, a finite element method-based software that computes physical data for electromagnetic and electromechanical phenomena. /// Autodesk of San Rafael, Calif., has released its Autodesk Algor Simulation 2010 products. /// A provider of CAD and CAM software for mold, tool, and die makers as well as for manufacturers of discrete parts, Cimatron Ltd. of Givat Shmuel, Israel, is now shipping CimatronE 9.0, which includes enhancements that enable faster creation of electrodes as well as new analysis tools. /// Elysium of Southfield, Mich., has released CADdoctor EX 4.1 for data translation. /// The developer of digital manufacturing applications using the XVL format, Lattice Technology of San Francisco, has released Lattice3D Dataway, which delivers multiple 3-D CAD data files directly into Lattice Technology’s XVL applications. /// Z Corp. of Burlington, Mass., has introduced its ZScanner 600, a handheld, self-positioning 3-D scanner. /// SmartDraw.com of San Diego, Calif., maker of SmartDraw software, has designed its product to work with Microsoft Visio software. /// Cranes Software of Bangalore, India, has released the latest version of its finite element analysis software, NISA Version 17. /// A maker of product data collaboration solutions, Actify Inc. of San Francisco, is now shipping SpinFire 8.4, which includes updates to SpinFire Reader, SpinFire Professional, and the introduction of SpinFire Markup and SpinFire Lite. /// Concepts NREC of White River Junction, Vt., has released Agile Engineering Design System 2009 for turbomachinery design and analysis.
|
