In the Loop: Ken Koedinger

Ken Koedinger is a professor of Human-Computer Interaction and Psychology at Carnegie Mellon University. A graduate of the University of Wisconsin at Madison with a B.S. in math and computer science and an M.S. in computer science, he earned his Ph.D. in cognitive psychology at CMU in 1990. Koedinger is a co-founder of Carnegie Learning Inc. and leads LearnLab, the Pittsburgh Science of Learning Center. He is the creator of Cognitive Models—computer simulations of student thinking and learning that are used to guide the design of educational materials, practices and technologies. Cognitive Models provide the basis for an approach to educational technology called Cognitive Tutors that support learning within rich problem-solving environments. He spoke to Link Editor Jason Togyer.

Where did you grow up?

Mostly the Milwaukee area of Wisconsin—where Herb Simon grew up.

What was your first computing experience?

In high school we had a programmable calculator—I’m sure my iPhone is thousands of times more powerful than it was. Later, as an undergrad at the University of Wisconsin, I got into environmental and biological system modeling on the Apple II. That was fun because you could access the memory of the Apple II directly with “peek” and “poke” commands—remember those? Other students in my class also had made computer models, and I built a generic simulation that could dip into any one of their models and modify the variables from a graphical user interface, without modifying their code.

What excited you about technology growing up?

I was captivated by the questions of “How do people do smart things?” and “How do people solve complicated problems and learn flexibly?” Those are great scientific problems, and computer science—through artificial intelligence—was providing a way to model them. Math and science came easy to me, but at the same time, I was interested in the social sciences, and I wanted to do something that would make a difference. Education is at the root of a lot of human problems. Dig into any social problem—war, poverty, hunger—and ultimately, you find that if people were better educated, they would be much easier to address. So I also was taking a lot of philosophy and psychology courses to study how people learn.

What attracted you to Carnegie Mellon?

I was very impressed with what was happening in computer science with Herb Simon, and in psychology with John Anderson, who became my mentor. John was doing basic science on how people think and learn while also working on applications. He was working on intelligent tutors that could teach LISP programming and high school geometry, and that was a perfect fit for my interests, because it combined understanding intelligence with doing something good in the world.

What about this process of “thinking” is so interesting?

We are far from cracking the mystery of how people learn, and why they’re so flexible and powerful in their thinking. The idea that Newell and Simon had—that if you have a good theory about thinking, you should be able to build a program that can replicate it—remains a big challenge. The “age of the machines” may be coming, but it’s going to be 20 or even 30 years before we have to worry about Skynet. Statistical methods have shown us how to learn categories and classify things, and made reasonable progress in natural language processing, but we’re not yet at the stage where systems can learn how to reason on their own.

Artificial intelligence has moved from using rules-based models to statistical models. Is there a shift back toward rules-based models?

I’m not sure that a “shift back” is the right way to say it. There are still some real mysteries that will require us to go back and revisit more symbolic, traditional approaches that Newell and Simon started with, but it won’t be instead of using machine learning. They’ll be a hybrid. The human mind is a statistical engine that soaks up patterns, but it also possesses conscious perception of the world, which we call reason. We don’t really understand yet the interface between the two, and we don’t have computer systems that can replicate it.

There were plenty of computer programs that were designed to teach skills. What made cognitive tutors so much more powerful than those?

The old message of psychology, going back to Freud, is that you don’t really know yourself—that so much is going on beneath the surface in the subconscious. We don’t know what we know. The key to cognitive tutors is that we start out by discovering, through data collection, what human experts really know. They can’t tell us. The two separate and equally challenging questions are, what it is that you’re trying to get students to do? And what teaching strategies work with the human brain to get them there?

What are some things we thought we knew about teaching, but which aren’t really true?

Well, algebra teachers for years have said students have the most trouble with story problems. We did a study comparing students who solved story problems with those who solved the same problems as matched equations, and we were really surprised—and so were math teachers—to find students did better with story problems than with matched equations. It reflected a lack of conscious understanding by teachers that they didn’t realize how much they knew about equations. It’s not just because they forgot—it’s because their brains have been working on the problem, subconsciously, over time. They’ve learned the language of algebra the same way you learn any other language.

We’re in the 10th year of the Pittsburgh Science of Learning Center. What have we learned?

The traditional approach to understanding what students need to learn involved interviewing experts, and was very labor-intensive. We’ve learned that we can use machine learning methods and data collection to make that a more efficient, scalable process. We now have more than 500 different data sets of people using intelligent tutors to learn language, science and math from kindergarten through college. The data repository and tools we’ve developed are helping researchers and course developers discover how students think and learn.

How will this sort of data inform the human teaching model 100 years from now?

We’re going to see less lecturing and more apprentice-style “learn-by-doing.” The reason we don’t have more learn-by-doing already is that there’s a scalability issue. Massive open online courses are very powerful and scalable because they’re focused on verbally delivering content-knowledge through lectures. But a lecture only gets at 30 percent of what you need to know to become an expert in a topic. To learn by doing, you need to be able to practice with feedback. You can get indirect feedback by doing homework and seeing the teacher’s comments the next day, but it’s not very efficient. If you have a one-on-one tutor, whether it’s a person or a computer, you can let your brain do its powerful soaking up of knowledge through practice, but you can also be guided and get some instruction when you’re stuck.

You taught in an urban high school. How was that experience?

I taught in Langley High School here in Pittsburgh as I was finishing my thesis. It wasn’t a very long experience, but it was incredibly eye opening, and my respect for teachers and the teaching profession—which was already high—continued to grow. It’s incredibly hard. Classroom management alone is very challenging, beyond the instructional challenges we’ve been discussing.

What do you like about teaching?

You see a student struggling, you figure out how to help them, and then you help them succeed. The satisfaction of programming is similar, right? You say, “I can’t get this computer to do this task.” But eventually it works, and it’s very satisfying. Teaching is a much harder task, but it’s also about seeing a student go from struggling to achieving. That turnaround is very satisfying.

What do you do for fun?

I work out on my elliptical. That’s fun—in an odd sort of way. I also play the guitar, and I have two daughters who are 6 and 2, and the 6-year-old is learning to play the piano, so my fascination with music is continuing.

For More Information: 

Jason Togyer | 412-268-8721 | jt3y@cs.cmu.edu