Ken Forbus

Ken Forbus's work in artificial intelligence is unique: He attempts to re-create the whole mind, not just a portion of it.


Justine Cassell

Justine Cassell with Sam, one of her Embodied Conversational Agents


Selim Shahriar

Selim Shahriar in his laboratory for atomic and photonic technologies


EECS research
Interdisciplinary reach, broad impact

Ken Forbus: Modeling the mind

Ask Ken Forbus, Walter P. Murphy Professor of Electrical Engineering and Computer Science, about his research, and his response is succinct: "We try to understand minds by building them."

Forbus has spent his career working in the area of artificial intelligence, attempting to understand how the human mind works by creating computer programs and simulations. His work is rare in that it tries to understand and re-create the whole mind, not just a portion of it. Most artificial intelligence systems have focused on aspects of cognition, such as skill learning. "That's a way to make a lot of progress, but skill learning is just a small piece of cognition," Forbus explains. "Hamsters get better at skills, but people have conceptual knowledge. They use that conceptual knowledge in all sorts of ways to deeply learn all sorts of things."

In order to create realistic models of how the mind works, Forbus partners with researchers in cognitive psychology. "There are incredibly valuable insights to be gained from psychology, especially cognitive psychology and other areas of cognitive science," Forbus says. For over 25 years he has worked with Dedre Gentner, professor of psychology in the Judd A. and Marjorie Weinberg School of Arts and Sciences, who also happens to be his wife. Their theories arise from the interplay of artificial intelligence and psychology, with Gentner doing experiments on people and Forbus building simulations as methods of testing their ideas.

The combination of computer simulation experiments and human psychological experiments provides more insights than either method alone. Each method has different strengths and weaknesses, giving researchers converging evidence for theories. Unlike traditional human psychological experiments, computer simulations provide the resources to test theories with a known set of variables. "We know what's in the memory of our computer simulation," Forbus says. "It's not the same as a human participant in an experiment. To be able to create something and know that these processes and these inputs will yield certain results is very powerful."

Forbus is currently working on "Companions Cognitive Systems," interactive systems with the ability for human-like learning. Central to Companions is the claim, from Structure Mapping Theory, that much of human learning is based on analogy. When presented with a new problem or situation, the mind identifies similar past experiences to form an action. A key element in this process is the ability to transfer knowledge from one domain, such as a subject area or skill set, to another. For example, if a person learns a lot about hydraulics, then parts of electricity should be easier to understand, since there are close analogies between the two subjects. Until now this type of learning was even more difficult for computers than for humans.

Forbus is working with the Educational Testing Service (ETS), the company that develops Advanced Placement (AP) and other standardized tests, to test Companions Systems' ability to learn AP Physics. ETS conducted an independent evaluation of a Companion, teaching it a small subset of AP Physics by giving it problems and worked solutions. "As you might guess, ETS is very good at generating problems," Forbus says. "There are different levels of transfer that have been identified: Can the system solve a problem it's seen before? Can it solve a problem with minor numerical variations? What if the numerical variations are so big that they actually change what happens?"

Forbus and his team started by giving a Companion a foundation of strategies and algebra skills but no knowledge of the equations of physics or when they should be used. When ETS would give the Companion a new type of problem, it would at first fail. But after giving it the worked solution, the Companion typically was able to solve many variations of the questions.

"When it sees a new problem, it figures out what would be a good analogy to something it has seen before, then tries to solve it. It solves problems very quickly," Forbus says.

In addition to being tested on AP Physics, Companions Systems has been challenged with problems drawn from tactical military games and strategy games, such as the popular open-source game FreeCiv. The initial success of these experiments provides hope for systems capable of general learning.

"We're trying to make a whole system that can learn about the domains it's working in, learn about the people it's interacting with, and learn about itself," Forbus says. Through his research, future computers may truly have a mind of their own.

Justine Cassell: Improving communication with and through computers

It's a common complaint: You call a company, only to be put through an endless series of computerized options. Your level of frustration continues to grow as you punch numbers and even tell an automated voice exactly what you want.

With all the frustration with computer communication, it's hard to imagine a future where computers might actually teach people a thing or two about communicating. Yet Justine Cassell, professor of communication studies and electrical engineering and computer science, hopes to do just that, not only improving how we communicate with computers but also using them to help people strengthen their ability to communicate with one another.

Cassell spent the early part of her career studying how people communicate with one another and since then has spent significant time studying how people interact with technology and use technology to interact with each other. Focusing on understanding natural forms of communication, Cassell works to develop technological methods to facilitate human-to-computer and human-to-human interaction.

"I think that for some time we've moved away from the metaphor of the computer as a tool toward the metaphor of the computer as a communication partner," she says. "My approach is, if the computer is our communication partner, it should know how to hold up its end of the conversation."

Cassell's work is unique in that it combines her training in linguistics and psychology with her interest in computer science. Using her understanding of human interaction, Cassell has developed a variety of "Embodied Conversational Agents," virtual humans able to interact with their human counterparts. The work combines verbal and nonverbal communication to create systems with a wide variety of potential uses.

"Many people who build technology start by building technology. I start with human-to-human behavior when building human-to-computer interaction," Cassell explains. "I try to think of something that we don't know about human-to-human communication, yet we assume when building human-to-computer interaction."

Cassell is the graduate director for a new PhD program in technology and social behavior, which incorporates faculty from McCormick and the School of Communication. Her research group comprises students who study both communication and computer science. It has developed a variety of virtual peers, including several for children and an interactive kiosk called NUMACK that gives directions around the Northwestern campus. The NUMACK research group studied how humans use gestures in providing directions and talking about campus. By closely examining the use of gestures and speech in interacting, the group is able to build computer models that can generate appropriate verbal and nonverbal communication when engaged with a human user.

Some of the virtual peers that the group developed act as learning partners to children. For instance, the virtual peer Alex helps children who speak African American Vernacular English (AAVE) at home to transition to Standard American English for the school environment. "Children who grow up speaking AAVE often have problems with literacy, which leads to a huge achievement gap between African American and white children in school," Cassell says. "Part of the issue is that teachers may not understand that AAVE is a valid dialect of English, so they think the child is speaking incorrect English. With all good intentions, they tell the child to stop speaking incorrect English, but the children don't yet know the alternative. For success in school, they need to do code switching speak AAVE at home and Standard American English in the classroom." Children build code-switching skills through interactive storytelling with the virtual peer.

Another virtual peer works with children with high-functioning autism, including Asperger's syndrome. This research group hopes that interaction with a virtual peer that can help develop both verbal and nonverbal communication skills, boosting a child's comfort level and leading to enhanced social activity and hence to increased learning in the classroom.

Cassell's development process, which starts with a principle of human communication and then uses computer science to re-create and test it, is a reflection of her own educational path. She holds master's degrees in literature and linguistics and a double PhD in linguistics and psychology not a typical résumé in computer science. Her work with virtual humans began out of frustration with the lack of tools needed to evaluate her hypotheses on human-to-human communication. She applied for a grant to spend a year in a computer science department to build a virtual human a project that shaped the course of her career.

In additional to developing virtual peers, Cassell has done extensive studies on how people interact with each other through technology. Long before Facebook, MySpace, and Friendster had entered the cultural lexicon, Cassell led the Junior Summit, an online kids-only community of 3,062 young people from around the world. She studied how young people formed community and chose leaders. Her work has made her an often-cited expert in the rapidly growing and highly discussed field of online social networking.

Cassell hopes her technological developments will have a broad impact on the way that humans use computers to communicate. "I hope that my work will make the computer accessible to a broader array of people those who may not be just like us, who may not communicate through typing, who may not be literate, and who may have communication issues," she says. "This kind of technology can act as a stepping-stone to full communication equality."

Selim Shahriar: Slowing down light to speed up communication

The speed of light is one of the few scientific parameters of which nearly everyone is aware. Light moves fast 3x108 meters per second, rapid enough to travel from the sun to the earth in just under 8.5 minutes. But the speed for which it is known is also one of the biggest obstacles to putting light to widespread use in practical applications such as high data-rate communication systems.

Selim Shahriar, associate professor of electrical engineering and computer science, is on the leading edge of answering the fundamental question of whether we can harness light. A significant part of his work focuses on "slow light," the technology that allows scientists to control light by slowing it down to as few as several centimeters per second. This research holds the promise for a wide variety of technological advances, including all-optical systems that could drastically increase communication data rates.

Most of today's communications systems use low-frequency electromagnetic signals transferred through copper wire during part of the communication process. These signals can be stored and rerouted to their final destination using common computer technology. But current systems have fundamental data-rate limitations and are quickly reaching their maximum capacities. On the information superhighway, they are the two-lane roads, with frequent bottlenecks.

All-optical systems systems that replace low-frequency signals with optical information could provide significant advantages. One major benefit is the ability to send signals on a large number of distinct channels through a single conduit, where each channel has more capacity than that of a copper wire carrying the low-frequency signals.

"In an optical fiber you can transmit light of many different colors simultaneously," explains Shahriar. "As long as they are slightly different in color, you could send a thousand or more kinds of light through one line, providing a huge increase in capacity."

Optical cables present a problem, however, even while providing incredible opportunities in terms of data rates. When a message arrives, it must be stored until it is determined where the message will ultimately land. In low-frequency systems carried by copper wires, digital information can be stored in computer memory, but current technology cannot store optical information as quickly as light can provide it.

This is where Shahriar's work in "slow light" comes in. He and other researchers studying slow light are developing ways to control and fine-tune the speed of light. The ability to adjust the speed of light provides the potential for broader use of optics in communication technology. The technology could allow for light to be slowed, rerouted, and delivered to an end user while maximizing the data rate.

Light slows down when traveling through a medium. For example, glass slows down light by a factor of about 1.5. However, the more light slows, the more it reflects off a surface, thus reducing the transmission. Most media can slow it only by a small factor (less than 10) before light reflects too much for a significant amount to pass through. This is hardly enough to provide real control.

One of the first breakthroughs in the field came when scientists were able to slow light to 17 meters per second with a new technique, using a cloud of ultracold sodium atoms, that does not suffer from this limitation. But "while the principle is helpful, it would be impractical to make anything out of a cloud of ultracold atoms," Shahriar comments.

Shahriar and his research group made the next breakthrough when they developed the ability to slow light using a crystalline matter. Building on the previous experiments with clouds of atoms, Shahriar's research was the first to use a solid material to slow light significantly, marking a step toward practical application of the technology.

Research in the field continues to expand, and scientists seek new advances in controlling light, such as the use of inexpensive materials or fiber-optic cables. Fundamental challenges must be overcome before it is realistic to plan for widespread use. As with any up-and-coming field, the applications for this technology are still under exploration. "Slow light is in its infancy, so we have to explore the ways that it can be useful," Shahriar says. "Eventually it should have applications that make people's lives better."

Slow light has the potential to provide communication systems that would enable computer transfers and downloads at data rates far beyond today's standards. It also has many applications in quantum information processing, which can be used to encrypt confidential information such as bank transactions.

While Shahriar's group continues with pioneering research in slow light, its studies have found other interesting ways to manipulate light. Using the same theories used to slow light, but in reverse, Shahriar's group can create light that appears to propagate faster than the normal speed of light. Although this "fast light" at first appeared to have no use, Shahriar has uncovered an interesting and potentially significant application, using it to enhance the sensitivity of rotation sensors in airplanes and missiles, thus making inertial navigation more accurate. The same effect could also be applied to test fundamental laws of physics with new precision and perhaps provide new insight into how nature works.

Whether being slowed down or speeded up, light has enormous potential for practical applications and improvements. Research at McCormick aims to take us closer to harnessing its power.

Kyle Delaney