Back to November 2002 CRN Table of Contents

[Published originally in the November 2002 edition of Computing Research News, Vol. 14/No. 5, pp. 1, 8.]

By Ron Brachman and Zachary Lemnios

The impact of the Defense Advanced Research Projects Agency (DARPA) on computing over the past 40 years has been profound. Led by the visionary J.C.R. Licklider and his innovative successors in the Information Processing Techniques Office (IPTO), DARPA initiated work that ultimately put personal computers on millions of desktops and made the global Internet a reality. In fact, the original IPTO, which lasted from 1962 to 1985, was in large part responsible for establishing Computer Science as a field.

DARPA has recently re-energized IPTO (now the Information Processing Technology Office), and has rededicated its attention to modern Computer Science by looking both to its roots and to a dramatic vision of the future. Licklider imagined computers and humans working closely together in a form of symbiosis. The new, 21st century IPTO wants to realize this vision by giving computing systems unprecedented abilities to reason, to learn, to explain, and to reflect, in order to finally create systems able to cope robustly with unforeseen circumstances. IPTO's goal is to create a new generation of cognitive systems.

Mired in Moore's Law?

One benefit of such cognitive systems would be their help in extracting us from a corner into which our success seems to have painted us. The research that has helped the industry follow Moore's "Law" has created processors that are remarkably fast and small, and data storage capabilities that are vast and cheap. Unfortunately, these incredible developments have cut two ways. While today's computers are more powerful than ever, we have been lured by processing power and inexpensive memory into creating systems that are enormously large and complex. Many of today's systems are virtually impossible for humans to understand, use, or maintain.

Beyond the resulting maintenance problem, with the total lifetime cost of systems now heavily dominated by after-production costs, this complexity has also led to serious vulnerabilities. More complexity means greater opportunity for intruders. More elements mean more ways that things can go wrong; systems crash and software rots. And the training burden and level of expertise required to cope with systems both keep growing. In order to make our systems more reliable, more secure, and more understandable, and to continue making substantial contributions to society, we need to do something dramatically different.

The Promise of Cognitive Systems

IPTO is attacking this problem by driving a fundamental change in computing systems. By giving systems more cognitive capabilities, we believe we can make them more responsible for their own behavior and maintenance.

Ideally, in the next generation, a computer system will be cognizant of its role in a larger organization or team (and of the overarching goals of that team), capable of acting autonomously, and able to interact rationally with other systems and humans in real time. It will also be able to take care of itself in a self-aware and knowledgeable way. Ultimately, these new capabilities will be the basis for artificial systems that can respond as robustly to surprise as natural systems can.

A cognitive computer system should be able to learn from its experience, as well as by being advised. It should be able to explain what it was doing and why it was doing it, and to recover from mental blind alleys. It should be able to reflect on what goes wrong when an anomaly occurs, and anticipate such occurrences in the future. It should be able to reconfigure itself in response to environmental changes. And it should be able to be configured, maintained, and operated by non-experts. All of these potential improvements in system capability should help us make a serious dent in the maintenance and complexity problems we are facing.

In a nutshell, we want to transform computational systems from those that are simply reactive to those that are truly cognitive. Our ultimate goal is to create systems that know what they're doing.

Where We're Going

New research in cognitive systems has the potential to revolutionize the way we design, deploy, and depend on computing systems. A long-term research agenda might be structured in stages. For example, we might strive first to consider software systems that were in some measure self-aware. This kind of system could help in its own debugging, and might be extensible through a high-level, goal-oriented dialogue with its programmer. Next, we could imagine building cognitive networks that are able to understand their overall goals and capable of making adaptive, effective use of limited resources. Beyond that, we are interested in building autonomous, perceiving agents, which could explain their reasoning and engage in natural dialogues with human partners that would allow them to increase their functionality and performance over time. Finally, we want to build truly intelligent, multi-component systems whose overall operation would be more efficient and more easily extensible.

The initial focus of our office will be on "assistant" or "associate" systems. The idea is to create an artificial system that could be a persistent, long-term partner for a person; this associate system would share experiences with its user and learn from those experiences. By being cognizant of the experiences of the user, the assistant system could be more effective in its communication. One can imagine an artificial executive assistant that becomes more and more personalized to its user over time, or a commander's associate that would become a dedicated partner for a battlefield commander, helping to anticipate his or her needs and removing the burden of administrative overhead.

Our effort will of necessity be multi-disciplinary, and will need to draw on many aspects of Computer Science. Despite its high-level focus on cognition, it will need serious participation from the systems, networking, security, and software communities, among others. The notion of architecture will be important throughout this work--truly cognitive systems are likely to be complex combinations of reactive processes, more thoughtful, deliberative processes, and reflective processes that capture self-awareness and help make the system robust in the face of unforeseen circumstances. Core technology will include learning, knowledge representation, reasoning, communication, perception, and multi-agent systems.

Why Now?

Many of the goals of our Cognitive Systems initiative are familiar. What makes us think that we are in a substantially better position to accomplish them now than we were before? We see several key factors: 1) improvements in computer hardware will soon give us computational substrates with the size and power to match the computational capability of animal and perhaps human brains; 2) the "Decade of the Brain" has brought us unprecedented insights from neuroscience, giving us new models of how the human brain works; and 3) there have been numerous successful deployments of a wide variety of artificial intelligence technologies, ranging from autonomous control of deep space missions to pragmatic machine learning improvements in speech understanding and data mining applications. While none of these or other factors is individually definitive, we believe that the convergence of computing power, knowledge of the brain, and practical experience in deploying reasoning and learning technology is remarkable.

As in many research endeavors, there is significant risk in this kind of initiative, but there is also extraordinary opportunity at a time when dramatically different approaches are urgently needed.

A Challenge for Computing Researchers

Just as the original IPTO owed its success to the energetic and creative talent in the then emerging field of Computer Science, our office must rely on the brainpower and imagination of research teams across the country. While we have issued a Broad Agency Announcement (DARPA BAA 02-21) that lays out our overall vision for Cognitive Systems, that vision cannot be realized without the individual and collaborative breakthroughs that have been the hallmark of American ingenuity.

Put simply: we need your good ideas. We encourage you to use our BAA as a catalyst for breakthrough thinking that might dramatically advance the state of the art.

We also need new Program Managers to help us define the programs that will bring our vision to life (and to secure the necessary funding). We're looking for a few passionate visionaries who want to have an impact on a national scale on our country's security and defense.

Off and Running

The revitalized IPTO is off and running. While our plans are ambitious, we believe we have no choice but to try something dramatically different in computing, lest we become victims of the complexity we have helped create. If we succeed, computer systems will be able to do substantially more powerful things. They will become easier to build and use, and will last longer. They will become the cooperative and supportive kind of partners that our predecessors imagined, and their novel capabilities will open new possibilities for both humans and machines. We hope you will join us in making cognitive systems a reality.

Ron Brachman is Director and Zachary Lemnios is Deputy Director of IPTO at DARPA.
To learn more, contact Dr. Brachman at rbrachman@darpa.mil or Dr. Lemnios at zlemnios@darpa.mil