Humans are still the "gold standard" of intelligent systems. Although machines have surpassed our capabilities in many particular domains, such as solving calculus problems and finding the shortest routes through graphs, no artificial system even comes close to the breadth, flexibility, and integration of capabilities exhibited by the average human. Even in those domains where we generally regard machines as having attained human-equivalent capability, this remains true only so long as we narrowly limit the domain: while a machine can generally solve calculus problems that a human can't, only a human is capable of sorting out which calculus problems are worth solving, or what a game of chess might reveal about the opponent's personality.
This fact is an inspiration, and not a discouragement. Looking backward, the field has had great success already in attaining human-level capabilities in narrow domains and shedding light on particular aspects of cognition. At the same time, we see the emergence of a new frontier in human-level AI research, where the problems of breadth, flexibility, and integration are beginning to be tackled directly. This yields a prospect both for revolutionary change in the capabilities of machines and also for a synthesis of cognitive models towards a broader understanding of the nature of human cognition.
Human-level AI is unapologetic in its lofty goals, which tend to emerge from two dominant and intertwined motivations: better understanding of human intelligence and increasing the capabilities of machines. This field stands squarely at the intersection of cognitive science and computer science. Researchers both draw on the tools of computing and derive important design constraints from knowledge about human and animal intelligence. While these interests sometimes compete, and many human-level AI researchers have at times been left to feel orphaned, more often this duality produces fruitful results for both fields.
On the one hand, the construction of machine systems can aid our understanding of human intelligence by illuminating particular aspects of human cognition. When a cognitive model is realized as a computational process, it's forced to be fully instantiated, often exposing subtle errors, unexpected constraints, and computational challenges. This is a double-edged sword, because it's usually the case that many details of that instantiation can't be grounded in our knowledge of human cognition. Even so, the insights gained from working with an operational model can produce surprising insights and generate new targets for investigation by cognitive psychologists and neuroscientists.
On the other hand, human-like cognitive abilities would be extremely valuable in many applications. Here, the study and modeling of human capabilities contributes both candidate mechanisms and engineering challenges. The ease with which humans solve problems that have previously appeared intractable, such as learning the meaning of words or reasoning about the beliefs of others, is a gauntlet thrown down to us as engineers, and we often discover new algorithms and engineering principles through the effort to model human capabilities. Though particular models of human-level behavior may work too hard at being faithful to the original to find immediate application, the insights that they yield have caused pragmatic revolutions before (for example, mathematical solvers, CAD tools, relational databases), and will continue to do so.
The focus of this special issue of IEEE Intelligent Systems is on an emerging new frontier in human-level AI research, where we are beginning to see the convergence of a variety of different approaches toward the goals of human-like breadth, flexibility, and resilience for AI systems.
The research on this new frontier is partially driven by both the continued rapid advances in cognitive studies of human intelligence and by the continued Moore's law increase of computational resources. Just as important, however, are difficulties that have been uncovered in the course of recent large-scale efforts like statistical natural language processing, the construction of the Cyc common-sense knowledge base, and the continuing development of traditional cognitive architectures such as Soar and ACT-R.
We thus see that the base goals of breadth and flexibility seem to be driving human-level AI efforts toward addressing three key challenges:
• scaling to massive amounts of background knowledge and/or computational mechanisms,
• principles that enable multiple representations or independent cognitive systems to cooperate in reasoning, and
• the ability to tolerate incorrectness and contradictions in a system's knowledge or models of the world.
In this special issue, we present four papers at the forefront of the new frontier in human-level AI research. Each brings a different background and perspective on the subject, and hence a different technical approach.
We begin with "Applying Common Sense Using Dimensionality Reduction," where Catherine Havasi, Robert Speer, James Pustejovsky, and Henry Lieberman grapple with scaling and the problems of integrating large data sets. The authors present a dimensionality-reduced representation of semantic networks called AnalogySpace, which they apply to find patterns, smooth out noise, and predict new knowledge based on the hundreds of thousands of relations in the Open Mind Common Sense ConceptNet. They further extend this to blend ConceptNet together with ontological knowledge from WordNet, generating useful new knowledge despite the incompatibilities between the representations of these two massive data sets and their internal inconsistencies.
Ken Forbus, Matt Klenk, and Tom Hinrichs take a different approach to similar problems in the second article, "Companion Cognitive Systems: Design Goals and Lessons Learned So Far." Here the main concern is integration of knowledge across domains, and the authors are engaged in a research program to see how much can be accounted for by analogical reasoning alone. This necessarily engages them with problems of inconsistencies between models as well as the management of the millions or billions of facts that an intelligent system might incrementally acquire during a lifetime of experience.
In "Reference Resolution Challenges for an Intelligent Agent: The Need for Knowledge," Marjorie McShane presents a careful examination of the range of challenges that an intelligent system faces in determining what objects are referred to by linguistic utterances. Although humans handle all of these with ease, purely syntactic approaches to language lack the necessary information. The article then goes on to lay out a program for addressing these problems through the systematic incorporation of certain categories of knowledge.
Finally, Nicholas Cassimatis examines logical mechanisms for doing just such integration of disparate knowledge sources in "Flexible Inference with Structured Knowledge through Reasoned Unification." Operating within the model of a cognitive substrate, the article develops a mechanism called "reasoned unification" that fills in missing information in script and frame representations by posing questions about identity relations to one or more reasoning mechanisms. This mechanism provides a rational approach to how different cognitive capabilities can work together to interpret ambiguous, implied, and nonliteral references.
These four articles are united by a common thread: each addresses a problem in achieving human-like breadth of capability, and is thus led to engage with problems of scaling and integration. Most importantly, however, each tells a clear story of how progress on the specific problems the authors are working on today leads to progress on the larger investigation of human-level intelligence.
For more information on this or any other computing topic, please visit our Digital Library at www.computer.org/csdl
is a scientist with BBN Technologies. His research interests center on the engineering of robust adaptive systems, with a focus on problems of system integration for human-level intelligence and on problems of modeling and control for spatially-distributed systems like sensor networks, robotic swarms, and cells during morphogenesis. Beal received his PhD in electrical engineering and computer science from the Massachusetts Institute of Technology. He is a member of the IEEE, AAAI, ACM, and the Cognitive Science Society. Contact him at firstname.lastname@example.org.
Patrick H. Winston
is the Ford Professor of Artificial Intelligence and Computer Science at the Massachusetts Institute of Technology (MIT). His Genesis Group studies how vision, language, and motor faculties account for intelligence, integrating work from several allied fields, including not only AI, but also computer science, systems neuroscience, cognitive science, and linguistics. Winston received his PhD in Computer Science from MIT. He is a past president of AAAI and a member of the Naval Research Advisory Committee. Contact him at email@example.com.