 |
 |
 |
 |
| |
|||||
|
|
|
Home > News & Events > MITRE Publications > The MITRE Digest > | |||||||||||||||||||
Neuromorphic Computing: Teaching New Brains Old Tricks May 2009
Editor's Note: This article and others on emerging technology research at MITRE and around the world can be found in our new publication, Envision. Designing computers that think like humans is difficult because computers and humans process information in fundamentally different ways. While computers are good at serial processing (one calculation at a time), human brains process information in parallel fashion (many calculations at the same time). Moreover, brains—in contrast to computers—are remarkably good at learning new things. Researchers are working to build computers with brain-like properties by creating new kinds of algorithms and microprocessors that replicate the features of biological brain circuits. These neuromorphic systems could one day perform challenging tasks like visual target recognition, navigation, and other complex behaviors that exceed the capabilities of today's machines. Deep Blue Versus Old Yeller: Why Brains are Better You're smarter than you think you are. And so is your dog. In fact, many of the cognitive abilities that we biological organisms take for granted—from our speediness at spotting a familiar face in a crowded scene, to our knack for anticipating the behaviors of others, to the ease with which we translate abstract goals into concrete sequences of actions—have proven to be the most difficult to replicate in artificial systems. Deep Blue may be able to mop the board with most humans in a game of chess, but pit IBM's vaunted supercomputer against your average canine in a game of "fetch my slippers," and Old Yeller wins every time (assume for the sake of fairness that Blue, like Yeller, has four legs and teeth). Unfortunately for the supercomputer, most of the challenges faced on a daily basis by biological organisms like ourselves have more in common with fetching slippers than with calculating the optimal move in a board game. The real world demands that we make sense of, and interact adaptively with, environments that are continuous, uncertain, dynamic, high-dimensional, and noisy—in other words, nothing like a chess board. Moreover, our brains accomplish all of this with remarkable efficiency. In terms of size (about 1.5 liters) and power requirements (less than 20 watts), the performance specs of a typical human brain are orders of magnitude more impressive than those of their closest silicon competitor. The Brain as a Blueprint So how does one build a machine with brain-like properties, i.e., a neuromorphic machine? One obvious first step would be to look toward the brain itself as a source of design inspiration. Let's watch Old Yeller in action as he performs his slipper-fetching duties. First he must rely on a myriad of cognitive skills during the course of his task: episodic memory (where did I last see the slippers?); spatial navigation (how do I get there from the bedroom?); dynamic sensory-motor control (how do I leap over the couch while at the same time dodging the toddler who's trying to grab my tail?); and cognitive control (how do I resist getting distracted from my mission by that savory scrap of bacon on the floor?). Furthermore, not only must all these skills work together as part of a seamless cognitive system, but they all must be flexible. In other words, they need to work just as well whether the task is fetching slippers, retrieving a newspaper, or even locating and destroying an enemy target. Modern neuroscience has made impressive progress in deciphering how the brain accomplishes these and other cognitive functions. New experimental tools, including non-invasive imaging techniques like functional magnetic resonance imaging (fMRI), have greatly expanded the range of hypotheses that can be explored by neuroscientists. Moreover, in recent years the field has become increasingly cross-disciplinary, with many young neuroscientists possessing strong backgrounds in mathematical modeling. As a result, theoretical and computational neuroscience have emerged as prominent subdisciplines. This new emphasis on formal computational models of brain function has in turn provided a workable starting point for engineers who seek to build systems that are truly brainlike in function. From Wetware to Hardware Even if one could build a computational model consisting of 100 billion neurons (approximately the number found in the human brain), implementing that model on a typical computer—or even a supercomputer, for that matter—is a whole different story. Conventional CPUs, which rely on serial information processing, are inherently unsuited for modeling the type of distributed, massively parallel processing that occurs in the brain. One branch of neuromorphic computing research, called neuromorphic engineering, is focused in part on developing custom microprocessors that do mimic the properties of neurobiological circuits in silicon. However, such neuromorphic chips are currently limited by poor scalability. The numbers tell the story: in a biological brain, any given neuron may be connected to up to ten thousand or so of its neighbors via junctions call synapses. In order to build a neuromorphic electronic device that matches the size, functionality, and power requirements of a mammalian brain, revolutionary new approaches are needed. Recently, Konstantin Likharev and colleagues at SUNY Albany have proposed a new class of neuromorphic electronic devices that utilize a hybrid of standard and nanoscale components. Likharev, along with others pursuing similar methods within academia and industry, project that this approach will enable the creation of neuromorphic devices that scale to densities and energy requirements on par with those of mammalian brains. Future Impact Whether the "brain" in question resides in a biological organism or an artificial one, the benefits of being able to function adaptively within a complex environment are clear. And while MITRE is presently pursuing neuromorphic research on behalf of its government sponsors, the technology that arises from that research will certainly have civilian benefits as well. For example, the same neuromorphic learning algorithms that may enable a military ground robot to learn local terrain features could also endow commercial video game characters with the ability to learn and adapt to the tactics of human players. Likewise, the same neuromorphic vision systems that might one day assist intelligence analysts in deciphering satellite image data might also form the basis for smarter image-based internet search tools. And because the Department of Defense is far from being the only enterprise suffering from massive information glut, any neuromorphic data mining tools developed for military applications would be well poised to transition to business applications. The Road Ahead At present, a fully functional electronic brain remains years—most likely decades—away. Next-generation neuromorphic hardware has yet to be implemented outside of simulation. Likewise, many of the computational brain models being developed by neuroscientists have not yet been fully validated. However, it seems likely that neuromorphic approaches will form the basis of a new generation of assistive technologies: adaptive computer interfaces, intelligent tutoring software, personal devices that sense a user's mood and social context, prosthetic devices for the disabled that integrate directly with the nervous system, and even intelligent household robots that do all the little things—like retrieve a pair of slippers. —by Brad Minnery Related Information Articles and News
Technical Papers and Presentations
Websites |
||||||||||||||||||||
Page last updated: May 22, 2009 | Top of page |
Solutions That Make a Difference.® |
|
|
