Vol 41 No 2: Summer 2020 | Published: 2020-06-29
The Defense Advanced Research Project Agency’s (DARPA) mission is to make pivotal investments leading to research breakthroughs that support national security. DARPA AI programs have emphasized the need for machines to perceive and interact with the world around them; to frame problems and to arrive at solutions and decisions based on reasoning; to implement those decisions, perhaps through consultation with a human or another machine; to learn; to explain the rationale for decisions; to adhere to rules of ethical behavior defined for humans; to adapt to dynamic environments; and, to do all of this in real time. In short, DARPA has always been interested in AI frameworks that integrate AI and computer science technologies, and the application of those frameworks to DARPA-hard problems.
In their introductory article to this special issue, DARPA’s Impact on Artificial Intelligence, Scott Fouse, Stephen Cross, and Zachary J. Lapin describe the significant role that DARPA has played in the establishment of artificial intelligence, and introduce five articles that explore DARPA’s three waves of AI. Their introduction is augmented by contributions by former DARPA director Steven H. Walker and former DARPA director Arati Prabhakar.
Knowledge Representation and Reasoning
A fundamental goal of artificial intelligence research and development is the creation of machines that demonstrate what humans consider to be intelligent behavior. Effective knowledge representation and reasoning (KR&R) methods are a foundational requirement for intelligent machines. The development of these methods remains a rich and active area of artificial intelligence research in which advances have been motivated by many factors, including interest in new challenge problems, interest in more complex domains, shortcomings of current methods, improved computational support, increases in requirements to interact effectively with humans, and ongoing funding from Defense Advanced Research Projects Agency and other agencies.
The article by Richard Fikes and Tom Garvey, Knowledge Representation and Reasoning – A History of DARPA Leadership, highlights several decades of advances in KR&R, paying particular attention to research on planning and on the impact of DARPA’s support. Fikes and Garvey are joined by David Israel, a principal scientist in the Artificial Intelligence Center at SRI International, who provides his own brief commentary on KR&R.
Richard Fikes, a professor emeritus of computer science at Stanford University, was director of Stanford’s Knowledge Systems Laboratory where he led major projects for DARPA and other Federal Government agencies focused on developing large-scale distributed repositories of computer-interpretable knowledge, collaborative development of multiuse ontologies, enabling technology for the Semantic Web, reasoning methods applicable to large-scale knowledge bases, and knowledge-based technology for intelligence analysts. Fikes is best known as codeveloper of the STRIPS automatic planning system, KIF (Knowledge Interchange Format), the Ontolingua ontology representation language and Web-based ontology development environment, the OKBC (Open Knowledge Base Connectivity) API for knowledge servers, and IntelliCorp’s KEE system.
Coauthor Tom Garvey is a senior principal scientist emeritus in SRI International’s AI Center where he pioneered work in active, goal-driven computer vision. His research career spanned work in computer vision, planning, approximate reasoning, command and control, human-directed autonomy, and learning. As program manager and assistant office director for planning and command and control in DARPA’s Information Systems Office he was responsible for programs in command and control, logistics, intelligence, and crisis avoidance.
Human Language Technology
Human language technology (HLT) encompasses a wide array of speech and text processing capabilities. The Defense Advanced Research Projects Agency’s pioneering research on automatic transcription, translation, and content analysis were major artificial intelligence success stories that changed science fiction into social fact. During a 40-year period, 10 seminal DARPA programs produced breakthrough capabilities that were further improved and widely deployed in popular consumer products, as well as in many commercial, industrial, and governmental applications. DARPA produced the core enabling technologies by setting crisp, aggressive, and quantitative technical objectives; by providing strong multiyear funding; and by using DARPA’s Common Task Method that was powerful, efficient, and easy to administer. To achieve these breakthroughs, multidisciplinary academic and industrial research teams working in parallel took advantage of increasingly large and diverse sets of linguistic data and rapidly increasing computational power to develop and employ increasingly sophisticated forms of machine learning.
In Human Language Technology, authors Mark Liberman, and Charles Wayne describe the progression of technical advances underlying key successes and the seminal programs that produced them.
The article also includes a brief commentary on HLT by Turing Award winner Raj Reddy of Carnegie Mellon University, an early pioneer in AI and a former AAAI president. Reddy is joined by Microsoft Technical Fellow and Chief Speech Scientist Xuedong Huang, who is the company’s key person behind Microsoft’s spoken language processing technologies.
Mark Liberman, a professor in the departments of Linguistics and Computer and Information Science, has participated in DARPA’s Human Language Technology programs since the mid 1980s. His involvement in the open data movement began with the ACL Data Collection Initiative in the 1980s, and continued with the provision of shared data for DARPA and other HLT programs, and the founding of the Linguistic Data Consortium in 1992.
Coauthor Charles Wayne played central roles in government-sponsored efforts to develop effective Human Language Technology during a long career in the defense and intelligence communities. He served as a program manager at DARPA from 1988 to 1992 and again from 2001 to 2005.
Machine learning methods provide a way for artificial intelligence systems to learn from experience. DARPA’s role in this fundamental area of artificial intelligence is explored in Joshua Alspector’s and Thomas G. Dietterich’s article, DARPA’s Role in Machine Learning.
Their article describes four threads of machine learning research supported and guided by the Defense Advanced Research Projects Agency – probabilistic modeling for speech recognition, probabilistic relational models, the integration of multiple machine learning approaches into a task-specific system, and neural network technology. These threads illustrate the DARPA way of creating timely advances in a field.
The article also includes a brief commentary on KR&R by professor emeritus Dan Hammerstrom of Portland State University, who served as a program manager at DARPA from 2012–2016, and Stanford University professor and Calico Inc. president Daphne Koller.
Josh Alspector’s long and distinguished career includes the first direct measurement of the velocity of the neutrino, the development (while at Bell Labs and Bellcore) of the first neural network learning microchip, the design of adaptive antispam filters. In 2009, he initiated DARPA’s Deep Learning program.
Thomas Dietterich, an emeritus professor of computer science at Oregon State University and former AAAI president, is one of the cofounders of the field of machine learning and has served the community in many roles including executive editor of Machine Learning, cofounder of the Journal of Machine Learning Research, and founding president of the International Machine Learning Society.
Vision and Robotics
Vision and robotics have the well-defined goal of meeting or exceeding human-level capabilities in perception, locomotion, and manipulation. Not surprisingly, that is perhaps easier said than done. Beginning in the 1970s, the Defense Advanced Research Projects Agency started the ambitious Imaging Understanding (IU) program that would continue for more than 20 years. The IU program began with fundamental research and slowly evolved into a host of more applied efforts with specific systems goals. Robotics programs followed a similar arc as the early research-oriented programs generated capabilities from which practical systems could be built. A culmination of the vision and robotics research was the DARPA Grand Challenge, which turned the impossibility of a self-driving car into an imminent reality.
In their article, Vision and Robotics, Thomas M. Strat, Rama Chellappa, and Vishal M. Patel tell the story of how some of the modern-day technologies we enjoy today trace their evolution from research sponsored by DARPA over the last 40 years. Their story is augmented with brief commentary by robotics entrepreneur Rodney Brooks (Rethink Robotics) and computer vision pioneer Takeo Kanade (Carnegie Mellon University.)
Thomas M. Strat (CEO of DZYNE Technologies), served as program manager at DARPA from 1995 – 2006. While there, he managed the Image Understanding Program, and created a half-dozen other programs involving vision, robotics, and autonomy. as well as serving as deputy director of the first two Grand Challenges for Autonomous Ground Vehicles.
Vishal M. Patel (Johns Hopkins University) is an assistant professor working on signal processing, computer vision, and pattern recognition with applications in biometrics and imaging.
Rama Chellappa is a Distinguished University Professor and a Minta Martin Professor of Engineering in the Department of Electrical and Computer Engineering at the University of Maryland.
Integrated AI Systems
From Shakey the Robot to self-driving cars, from the personal computer to personal assistants on our phones, the Defense Advanced Research Projects Agency has led the development of integrated artificial intelligence systems for more than half a century. From the earliest days of AI, it was apparent that a robust, generally intelligent system should include a complete set of capabilities: perception, memory, reasoning, learning, planning, and action; and when DARPA initiated AI research in the 1960s, ambitious projects such as Shakey the Robot went after the complete package. As DARPA realized the challenges, they backed away from the ultimate goal of integrated AI and tried to make progress on the individual problems of image understanding, speech and language understanding, knowledge representation and reasoning, planning and decision aids, machine learning, and robotic manipulation. Yet, even as researchers struggled to make progress in these subdisciplines, DARPA periodically resurrected the challenge of integrated intelligent systems and pushed the community to try again.
In the 1980s, DARPA’s Strategic Computing Initiative took on challenges of integrated AI projects such as the Autonomous Land Vehicle and the Pilot’s Associate. These did not succeed, but instead set the stage for the several decades of more siloed research that followed, until it was time to try again. In the first decade of the twenty-first century, DARPA took on the integrated AI problem again with its Grand Challenges, which led to the first self-driving cars, and projects such as the Personalized Assistant that Learns, which produced Siri. These efforts created complex, richly-integrated systems that represented quantum leaps ahead in machine intelligence. The integration of sophisticated capabilities in a fundamental way is the key to general intelligence.
In Integrated AI Systems, authors Ron Brachman, Dave Gunning, and Murray Burke weave the story of DARPA’s persistent long-term support for this essential premise of artificial intelligence. Their story is augmented with brief contributions by Sebastian Thrun, who led development of the 2005 DARPA Grand Challenge winning robotic vehicle Stanley, and then went on to lead development of the Google self-driving car, and Adam Cheyer, chief architect of the CALO/PAL project and later cofounder of Siri Inc.
Former AAAI President Ron Brachman is the Director of the Jacobs Technion-Cornell Institute at Cornell Tech in New York City and a professor of computer science at Cornell University. He was the director of the Information Processing Technology Office at DARPA from 2002 to 2005. His distinguished research career has focused on knowledge representation and reasoning.
David Gunning has managed AI projects at DARPA over the last three decades as a program manager from 1994-2000, 2003-2008, and 2015-2019. His programs included the Command Post of the Future (CPOF), the Personalized Assistant that Learns (PAL), and Explainable AI (XAI).
Murray Burke was the program manager for DARPA’s High Performance Knowledge Bases, Rapid Knowledge Formation and DARPA Agent Markup Language programs. After completing his appointment at DARPA, Burke has been providing systems engineering and technical assistant (SETA) services to numerous DARPA AI programs over the past sixteen years including the Personalized Assistant that Learns and Deep Exploration and Filtering of Text programs.
DARPA's AI Vision
The special issue concludes with a short commentary on DARPA’s AI Vision by DARPA’s acting director Peter Highnam. He is joined in this commentary by AAAI President Yolanda Gil.