Artificial Intelligence Merges With Neuroscience-Will Intelligent Machines Follow?
Can intelligent, manmade machines someday replace soldiers on the battlefield? The US Department of Defense, with the help of a number of academic institutions and research companies, is hoping that recent advances in neuroscience and cognitive psychology, combined with the computer science and engineering principles of artificial intelligence, will ultimately lead to the creation of machines with real cognitive capabilities. The better our understanding of human intelligence, Pentagon officials believe, the better the odds of creating machines that can simulate human cognition.
The quest to turn science fiction into reality is the mission of the Biologically Inspired Cognitive Architectures (BICA) program, the brainchild of the Defense Advanced Research Projects Agency (DARPA), the Department of Defense’s central research and development organization. Officially, BICA was designed to develop improved computational models of human cognition that could be used to create improved artificial algorithms for machine perception, reasoning, learning, and intelligence. Faced with a number of automation challenges, which include creating autonomous systems that perform reliably without constant human intervention as well as advanced military intelligence, surveillance, and reconnaissance systems, DARPA and the Department of Defense will depend on the creation of more flexible, competent, autonomous, and cognitive software, according to David Gunning, MS.
“With the BICA program, we are looking to the best-known example of a cognitive system-the human brain-for inspiration and guidance,” Mr. Gunning told Neurology Reviews. “The BICA program will implement and evaluate psychologically and neurobiology-based architectures of human cognition that could be used to guide creation of more human-like cognitive capabilities in machines.” Mr. Gunning is a program manager at DARPA’s Information Processing Technology Office in Arlington, Virginia.
In its nearly 50-year-history, DARPA has been the impetus behind the Saturn V rocket, stealth technology, the Internet, and the field of unmanned aerial, underwater, and ground vehicles. “If it’s not revolutionary change with radical new ideas, we are not interested in it,” said Brett P. Giroir, MD, DARPA’s Deputy Director of the Defense Sciences Office, at the recent 35th Critical Care Congress in San Francisco. “There is now a strategic thrust at DARPA called the biorevolution. DARPA has taken on the mission to explore those ‘crazy’ ideas in biology that can really lead to revolutionary improvements in our defense capabilities.”
2006: A Brain Odyssey
The long-term goal of the BICA project “is to finally solve the artificial intelligence problem,” Randall O’Reilly, PhD, told Neurology Reviews. “DARPA has realized that traditional symbolic approaches are just not going to get there, and that looking to our one known working example-the brain-is likely to provide important insights and some guarantee of success, at least in the long run. Once you have solved the artificial intelligence problem, virtually anything that humans do can be automated and done cheaper, better, faster, etc. Thus, the practical implications of this are likely to be major and widespread, transforming society and life as we know it. This is likely to take a long time and happen incrementally, so we’ll have plenty of time to ease into it.” Dr. O’Reilly, who heads one of the BICA research groups, is an Associate Professor of Psychology, Institute of Cognitive Science, Center for Neuroscience, University of Colorado, Boulder.
“We still can’t build systems even with the perceptual recognition capabilities of a mouse or a dog, let alone the higher cognitive abilities of human beings,” Richard Granger, PhD, commented to Neurology Reviews. “Procedures that seem easy and natural to humans-for example, language-and even to other animals-such as recognition of images, sounds, and odors-have been notoriously difficult for artificial systems to perform. Such tasks as speech understanding are specified only by our own abilities. Many current engineering systems, such as automated telephone operators, are pretty good. And the only reason that we know that these artificial systems can be outperformed is that humans outperform them. Research in the neurosciences over the past decade has generated vast amounts of data directly bearing on these questions.” Dr. Granger, Professor of Computer Science and Cognitive Science and Director of the Brain Engineering Laboratory at the University of California, Irvine, is working in conjunction with project team leader Anna Tsao, PhD, President of AlgoTek, Inc. in Annapolis, Maryland.
The most pragmatic approach to construct intelligent systems may begin with understanding the underlying principles of natural intelligence, offered Dr. Granger. “Just as flying machines were based on principles of aerodynamics used by flying animals, intelligent machines will arise from understanding the principles of intelligence,” he said. “The only instances of those principles are biological systems, and thus the goal of the program is to educe these biological principles and apply them to the construction of biologically inspired systems.”
Innovative Ideas
Phase I of the BICA program began in October 2005 and is expected to run through November 2006. After soliciting for proposals, DARPA selected those that were the “most technically competent and innovative,” noted Mr. Gunning. All involve developing theories of cognition, mapping cognitive functions to neurologic structures, or designing psychologically and neurobiologically based cognitive architectures. Teams include researchers from the Massachusetts Institute of Technology, University of Michigan, Carnegie Mellon University, University of Colorado, Hughes Research Laboratories, University of Maryland, AlgoTek, Inc., Harvard University, Rutgers University, Numenta, Inc., George Mason University, Klein Associates, Inc., Emory University, University of Pittsburgh, and University of Southern California.
“The phase I performers are, in some cases, refining and extending computation models they had previously developed, while others are starting from scratch to design novel architectures based on new discoveries in cognitive psychology and neuroscience,” said Mr. Gunning. DARPA hopes to begin soliciting this fall for teams in phase II to build and test more complete computational models of human cognition within four to five years.
Linking Sensation, Memory, and Action
Another group of investigators, led by Mark A. Gluck, PhD, is building a model of the hippocampus that will integrate the functionality of adaptive brain systems for sensation, learning, memory, and action. The resulting neurocognitive computational model would have the potential, for artificial intelligence and autonomous control purposes, to “explain and exploit advanced capabilities arising from brain systems that cooperatively realize our high-level cognitive capabilities for sensory processing in multiple modalities, pattern recognition, memory storage and retrieval, skill learning and generalization, temporal prediction, and adaptive real-time motor control,” according to Dr. Gluck. He is a Professor and Codirector of Rutgers University’s Memory Disorders Project, Center for Molecular and Behavioral Neuroscience, in Newark, New Jersey.
Overall, BICA’s aim, Dr. Gluck told Neurology Reviews, is “to harness the incredible progress made in the last 10 years in neuroscience and cognitive science to guide the development of a new generation of artificially intelligent computer systems that can replace soldiers in the battlefield in many situations. This is the natural complement to the unmanned plane and vehicle programs at DARPA. The goal is to have within four years a clearer sense of the basic principles and architectures for building brain-based computer systems to inform future applied systems for various specialized applications.”
Modeling the Brain
Dr. O’Reilly’s group is concentrating their research on the posterior cortex, hippocampus, and prefrontal cortex and have spent the past 15 years developing increasingly sophisticated computer models of these brain areas. “I believe that we now have a reasonable consensus theory of the fundamental operations performed by these areas and the neural mechanisms that support them,” said Dr. O’Reilly. “Our implemented models can perform many of the key, defining cognitive tasks associated with each of these areas. For example, our posterior cortical models can do sophisticated object recognition-such as recognizing 100 everyday objects regardless of considerable variation in size, rotation, and retinal position; spatial processing, including transforming between different coordinate systems; categorization; associative memory; semantic processing in language; etc. Our hippocampal models can do many different forms of episodic memory tasks (recognition, recall, etc). Our prefrontal cortex models can perform benchmark tasks like the Wisconsin Card Sorting Task, Stroop, “AX”-type continuous performance task, etc. These are the tasks that behavioral researchers have identified as depending critically on the unique mechanisms of the prefrontal cortex.” The researchers are now putting these brain area models together into an integrated tripartite cognitive architecture, with plans to move from the lab to the real world. “We have a plan that should end up producing a fairly compelling intelligent agent, though still somewhat simple relative to humans,” he said.
Cognitive Architecture
Dr. Granger’s team is focusing on the roles that various parts of the brain have regarding perception, memory, and speech production. As neuroscientific advances have yielded new insight regarding human brain circuitry, “Some of the resulting derived computational algorithms have been surprising, encompassing abilities that were not thought to emerge directly from brain circuit operation,” noted Dr. Granger. “Taken together, the operations of individual circuits interact to give rise to larger routines that are hypothesized to comprise the instruction set of the brain-that is, the basic mental procedures from which complex behavioral and cognitive operations are assembled. These findings give rise to specific hypotheses about what kinds of biological ‘steps’ occur during particular mental operations. These have led to biological and psychological predictions that are being tested experimentally and to new approaches that may help lead to novel solutions to complex computational and engineering problems.”
Hurdles and Obstacles
There is a consensus that much work needs to be done before researchers are able to build machines that can process visual images and respond to them in an intelligent manner. And how one defines the word intelligent in this regard is a matter of debate, commented Dr. O’Reilly. “We and others have developed object recognition models that are increasingly robust and should soon be able to get past those ‘distorted letter code’ tests that you find on Web sites to prevent automated Web-bots from going further,” he said. “Intelligence really is a graded construct, but I think the brain-based models are already beyond the kinds of rigid, rote mechanisms that many people are familiar with. Within a couple of years, these systems will likely enter the commercial marketplace, and progress will continue from there.” Added Mr. Gunning, “As in any research endeavor, it is very difficult to estimate what you don’t know and what obstacles you have yet to overcome. We know that we cannot build computer systems that do this today. We are very confident that we will be able to do better than we do today, but it is very difficult to estimate when we will achieve dramatic improvements to computer perception and intelligent responses.”
Will Machines Think as Well as Humans?
Mr. Gunning believes that the ultimate step of creating computers with full human intelligence will not happen anytime soon. “We are very far from creating a computer that genuinely thinks as a human,” he said. “However, I am also very confident that within five years we will be able to create computer models of cognition that are dramatically more like human cognition than the ones we have today, and ones that could revolutionize our ability to build software systems that can reason, learn, and adapt to changing situations. I want to emphasize that our goal in the BICA program is not to create general computer systems that think like humans, but rather to create systems that emulate human processing to provide much more efficient and capable processing for specific applications. Such systems would be highly skilled at the applications they were intended to address and would be able to acquire knowledge in one situation and use it in other similar situations, but would in no way have the broad, general knowledge and capabilities that humans exhibit.” Speculating on a timetable for creating smart machines is “highly uncertain, because we just don’t know enough to give good estimates,” added Dr. O’Reilly. “Pessimists would say hundreds of years. I’m an optimist, and I see that we already have many small-scale but quite functional models of many key brain areas. So I think that we will make major strides in the next five to 10 years just by gluing these existing models together and challenging them with increasingly complex cognitive tasks. One of the major limitations is computational power, but this should be increasing at about the rate that is needed to keep pace with the developing scientific understanding of how the brain works. At the very least, after five to 10 years from now, I think we’ll have a much better estimate of how long it will really take.”
Before having any chance to create a machine that can think as well as a human, investigators must fully learn the scientific principles by which the human brain achieves intelligence, Dr. Granger pointed out. “The brain is arguably the most complex object known to science,” he said. “The task is very large and requires research that cuts across traditional disciplinary boundaries. Crucial information will come from neuroscience, psychology, computer science, mathematics, and other related fields. It will take a consortium of researchers from multiple fields and multiple institutions to tackle this collaborative endeavor. It is impossible to know, at this early stage, how far we are from creating artificial cognitive systems. But a solid scientific understanding of the brain is the road that leads to them, and we are walking that road.”
-Colby Stong