The Next Frontier of Learning Engineering: AI That Teaches Other AI

This post was originally published on this site

Humans tutoring other humans works pretty well. The trouble is, it requires a lot of people. Artificially intelligent tools tutoring humans works pretty well, too—but building those digital systems takes time and expertise.

So researchers hoping to engineer better teaching and learning systems are working to unlock a new level of education efficiency by creating AI tools that make it easier for almost anyone to build an AI tutor.

“We are trying to leverage the joint power of human tutoring and computer tutoring,” says Ken Koedinger, a professor of human-computer interaction and psychology at Carnegie Mellon University.

Speedier Solutions to Answering Student Questions

Creating the kind of AI tutoring tool that complements or even replaces the work of a human tutor can take skilled computer programmers hundreds or thousands of hours. That puts such tools out of reach for most teachers looking for new ways to provide their students with personalized support.

“No teacher is going to put in 1,000 person-hours of his or her time in order to get a benefit of 200 person-hours that he or she may save,” Ashok Goel, a professor of computer science and cognitive science at Georgia Institute of Technology, told EdSurge in an interview earlier this year. “It’s not something I could hand over to you or to some colleague and say, ‘Go run it in your class.’”

A thousand hours is about how long it took Goel and his team to create Jill Watson, an AI teaching assistant chatbot that can answer student questions. Now, Goel and his colleagues are working on a new tool that can build a Jill Watson with just a bit of human help. Called Agent Smith, it’s an artificially intelligent system that absorbs information from a course syllabus and uses it to build a Jill Watson customized to that class. Doing so takes only about 10 hours of work from a human.

The power to produce AI education tools in a fraction of the original time is exciting to Goel, who thinks every teacher, child and parent should have access to a Jill Watson.

“I think it’s doable now that we have Agent Smith,” he says. “If we can do it in two hours—we’re not there yet, but if we can do it in two hours, then I can see the scaling up really happen.”

Minting More Math Tutors

Meanwhile, Koedinger and other researchers at Carnegie Mellon University are working to create a system that can easily learn math skills from a human teacher, and then tutor students in those skills.

When the first AI tool in the system, called the Apprentice Learner, encounters a new type of math problem, it will ask a human user to demonstrate a step-by-step solution. The Apprentice Learner then makes hypotheses about how the solution steps work and tests those theories on subsequent problems. The human user offers positive or negative feedback from which the tool learns. See the tool in action here.

Building a tool that learns the same way a student learns—through practice and feedback—means that “a nonprogrammer now can essentially teach the computer by demonstrating,” Koedinger says. And it can also yield insight about what makes learning hard for humans, he adds, because when the Apprentice Learner struggles, “it’s pretty predictive of when a real student is going to struggle, often in ways human experts don’t realize.”

In turn, the Apprentice Learner uses what it knows to create intelligent tutoring systems that offer that same kind of math practice and feedback to human students.

“The teacher teaches one ‘student,’ and the computer teaches all the rest,” Koedinger says. “The code is getting written by artificial intelligence.”

The researchers would like to improve this system such that teaching it a new skill takes a human educator the same amount of time as it takes to tutor a student directly.

“Even faster would be great,” says Carnegie Mellon doctoral student Daniel Weitekamp. “There are still a few bugs, but we’re rapidly getting there.”

And because teachers often prefer differing strategies for solving math problems, the system can learn alternative solution paths to suit a variety of methods.

“One teacher can make their tutor strict. Another can make it more flexible,” Koedinger says. “You can do it your way. It opens up more doors.”

Building Better Online Courses

Slashing the time it takes to create an entire online course—one that incorporates artificially intelligent personalized tutoring—is the goal of Korbit, a Canandian startup founded by alumni of Montreal’s Mila artificial intelligence research institute and Cambridge University.

Online education tends to be widely accessible, but online course completion rates are low. AI tutors can boost student learning, but they’re resource-intensive to make. Korbit aims to combine the best of both education systems without all the human labor that typically goes into creating either.

“It takes a really long time to build these programs—a year and a team of 10 people to build one physics course,” says Iulian Vlad Serban, CEO of Korbit. “There are lots of issues, and the biggest one [is] scalability.”

The company is working on AI technology—called Korbi—that reduces the time it takes to create effective, interactive online courses that include chatbot-based tutoring supports such as hints and definitions. It’s based on “an algorithm that sits on top of other algorithms,” Serban says.

Teachers construct the building blocks—the course modules—and Korbi organizes them for students according to their personal goals and what lessons the tool discerns they need. For the tutoring component, the tool draws on data that teachers put in the system—and from information it gathers from Wikipedia and open educational resources.

“We don’t write a thousand rules,” Serban says. “A teacher writes the questions, writes one or two answers. Korbi analyzes that, and scrapes data from the web and builds out the course.”

Pulling information from the internet hasn’t resulted in a lot of inaccuracies so far, he adds, but it does sometimes pull in irrelevant facts.

“The main problem we are working on is finding the most relevant piece of information the student needs,” Serban says.

Korbi is as much a student as it is a tutor. Over time, the system adapts the interventions it offers human users as it learns what works. The fact that the tool can teach at scale, for thousands of people at once, also means it has access to large quantities of the information it uses to improve.

“We let the AI algorithm figure it out from its own data,” Serban says. “Most of what it does is learning from the students. Students are teaching it to do better.”

%d bloggers like this: