DARPA awards $2.8 million to Cognitive Computing Lab

The Cognitive Computing Lab (CCL) at Tech recently earned $2.8 million in funding from the Defense Advanced Research Projects Agency (DARPA) to work on cognitive computing projects.

“Cognitive computing is a blend of AI [Artificial Intelligence] and HCC [Human-Centered Computing], where we study and build intelligent systems that can function in environments with other cognitive beings,” said director of the CCL Ashwin Ram.

“You can look at Deep Blue, which is an amazing piece of software and a very good chess player, but there are no cognitive principles involved in its construction. It does not help us understand any better how humans play chess,” Ram said. “On the other hand, there is a lot of work in the cognitive sciences that has not transferred to computing systems. For example, much is known about human communication but we do not have systems that can communicate in English like a person does. Cognitive computing makes the connection in this sense between humans and machines.”

This effort is part of a four-year, $25 million project on Generalized Integrated Learning Architectures (GILA), led by Lockheed Martin and involving academic and industry entities such as Arizona State University, the University of Massachusetts, Fujistsu Labs and the Georgia Tech Research Institute (GTRI).

“If you look at some traditional AI systems, they require a lot of training and are very specific to one task. The learning they do does not transfer very well. You can train a neural network, for example, to recognize your handwriting. When it is done it recognizes only your handwriting, not mine. It does not recognize your voice or your face,” Ram said. “GILA is intended to build a system capable of learning very quickly based on one or a few examples by using different learning techniques and integrating diverse kinds of knowledge.”

But what could a computer learn exactly?

Basic proven applications include the assembly of furniture, for example. However, more interesting types of knowledge have been identified in more complex and realistic domains such as real-time strategy games.

“We have a group of faculty working on game AI, which includes Charles Isbell, Michael Mateas and me. We are going to build an agent that can think strategically, develop plans for what it wants to do in the game, play the game and then learn from that experience,” Ram said. The lessons learned from this effort can also shed some light into the way the human mind works.

“We want to understand what makes us humans be able to learn so well and then duplicate this ability to get more human-like performance from machines,” Ram said.

Tech’s involvement in this project is very significant and larger than that of many of the other contributors. “We have a nice slice of the pie here,” Ram said. The team at the CCL has three main roles: to develop some of the techniques that the machine will use to play the game and solve problems, to develop what has been called the ‘meta-reasoner’ which combines all the techniques and allocates resources and to develop the game domain itself.

Michael Mateas, assistant professor at the CCL, is in charge of developing the language bridging the gap between the planning at the strategic level and the execution level in the game.

Collaboration between experts of different fields has proven vital in similar past projects and will certainly be paramount for the GILA project.

“We in the CCL are very interested in interdisciplinary research. We don’t want a traditional department that looks inward but rather one that looks outward to all the disciplines that it can support and interact with. I think this project is a good example of the kind of broad-base collaboration that the College of Computing (CoC) is all about,” Ram said. Departments that have collaborated in the past with the CoC include the Literature, Communication and Culture (LCC), psychology, industrial, electrical and mechanical engineering, as well as private industry and GTRI.

One of the four main goals of the program is to transition the Integrated Learning software for military applications to enable low-cost military decision/planning support systems. War-game simulators are used extensively in the training of military personnel. In principle more realistic technologies embedded with smarter agents result in superior training experience.

“One of the best pieces of educational software made is the flight simulator because it behaves just like a real airplane,” Ram said. “Currently a battlefield simulator doesn’t really work like a battlefield because its agents are not smart enough. We are trying to improve this with the GILA project.”

This technology, however, has a vast realm of applicability. One of them is an Air Tasking Order (ATO) planning system that learns human planning processes for air operations and imitates them as a decision support asset.

Air operations campaign planning is a complicated and time-sensitive process that involves a large number of people and software systems to plan the activities of thousands of aircraft, crews, support staff and support logistics.

Another application is the Delivery AtlasQuest—a system that learns to construct plans for truck deliveries in urban areas. The Intelligent Travel Assistant (ITA) is an application that learns complex travel plans by watching a person plan his or her own travel, involving a full spectrum of activities such as flight arrangements, car rental, dining reservations and sporting events.

There is also significant interest in the commercial field, more specifically in the entertainment gaming industry. “Games sell on two things now: better graphics and better AI,” Ram said.

A critical aspect of this technology, however, inevitably points back at Hollywood’s visions and presents the question of who is in control, or, in more practical terms, who is really making the decisions.

“One of the key things is for humans to be comfortable with the decisions the system is making, and in order to do that, the decisions have to be transparent,” Ram said.