Cutting Edge of AI – Teaching Its Clone
Humans are pretty good at tutoring other humans. The problem is, it involves a lot of individuals. Artificially advanced human tutoring platforms still work pretty well, but it takes time and experience to develop such automated systems.
So researchers hoping to engineer better teaching and learning systems are trying to unlock a new level of education productivity by developing AI resources that make it easier for almost everyone to create an AI tutor.
Ken Koedinger, a professor of human-computer interaction and psychology at Carnegie Mellon University, says, “We are trying to harness the combined strength of human tutoring and computer tutoring.
Faster Path to New Solutions with AI
It can take hundreds or thousands of hours for professional computer programmers to build the sort of AI tutoring tool that complements or even replaces the work of a human tutor. For most teachers searching for new ways to offer personalized assistance to their students, this places those resources out of control.
“In an interview earlier this year, Ashok Goel, a professor of computer science and cognitive science at the Georgia Institute of Technology, told EdSurge, “No teacher would put 1,000 person-hours of his or her time to get a gain of 200 person-hours that he or she will save. “It’s not something that I’ve been able to hand over to you or a colleague and say, ‘Go run it in class.'”
How long it took Goel and his team to build Jill Watson, an AI teaching assistant chatbot that can answer questions from students, is a thousand hours. Now Goel and his colleagues are working on a new tool that, with just a little human support, can create a Jill Watson. It’s an artificially intelligent device called Agent Smith, which absorbs knowledge from a course syllabus and uses it to create a personalized Jill Watson for that class. Doing so just takes a human being about 10 hours of work.
In a fraction of the original age, the power to create AI education resources is exciting for Goel, who thinks every teacher, child, and parent should have access to a Jill Watson.
“Now that we have Agent Smith, I think it’s doable,” he says. “We’re not there yet if we can do it in two hours, but if we can do it in two hours, then I can see the scaling up happening.”
Creating New Teachers from AI
In the meantime, Koedinger and other researchers at Carnegie Mellon University are working to build a device that can effectively teach a human teacher’s math skills, and then mentor students in those skills.
When a new form of math problem is encountered by the first AI instrument in the system, called the Apprentice Learner, it will ask a human user to show a step-by-step solution. The Apprentice Learner then hypothesizes how the steps of the solution work and checks certain hypotheses on subsequent issues. The human user generates constructive or negative feedback that the tool learns from. See the tool here in motion.
Building a tool that learns the same way a student learns means “a non-programmer can now essentially teach the computer by demonstrating,” Koedinger says, through practice and feedback. And it can also provide insight into what makes learning difficult for humans, he says, since “it is pretty predictive of when a real student will struggle, often in ways that human experts do not realize,” when the Apprentice Learner struggles.
The Apprentice Learner, in turn, uses what he learns to build intelligent tutoring programs that give human students the same kind of math practice and input.
Koedinger says, “The teacher teaches one student, and the computer teaches all the rest.” “Artificial intelligence will write the code.”
The researchers would like to refine this method in such a way that it takes the same amount of time for a human instructor to instruct a student directly to teach him a new ability.
“Even faster would be great,” says Daniel Weitekamp, a Carnegie Mellon doctoral student. “There are still a few bugs, but we will get there quickly.”
And since teachers also prefer various math problem-solving techniques, alternate solution paths can be taught by the system to accommodate a range of methods.
“An instructor should make the mentor strict. Another could make it more versatile,’ says Koedinger. Your way, you can do it. That opens even more doors.
Continuing Education Takes a Big Leap Forward
The aim of Korbit, a Canadian start-up created by alumni of Montreal’s Mila Artificial Intelligence Research Institute and Cambridge University, is to reduce the time it takes to build an entire online course, one that integrates artificially intelligent personalized tutoring.
Online education continues to be widely available, but completion rates of online courses are low. AI tutors can improve learning for students, but they are resource-intensive. Without all the human work that usually goes into producing one, Korbit seeks to incorporate the best of all education systems.
“Building these programs takes a long time, a year, and a team of 10 people to build one physics course,” says Iulian Vlad Serban, CEO of Korbit. “There are many problems, and the greatest one [is] scalability.”
The company works on Korbi-called AI technology, which reduces the time it takes to build effective, interactive online courses that provide support for chatbot-based tutorings, such as hints and definitions. Serban says it’s based on “an algorithm that sits on top of other algorithms.”
Teachers create the building blocks, the course modules, and Korbi organizes them according to their objectives for students and what lessons the tool discerns they need. The tool draws on data that teachers put into the framework for the tutoring portion, and from the information, it collects from Wikipedia and open educational tools.
“There are a thousand rules we do not write,” Serban says. The questions are written by an instructor, one or two answers are written. Korbi analyses this scrapes information from the web and designs the course.
So far, pulling data from the internet has not led to a lot of inaccuracies, he says, but it also pulls in irrelevant facts.
“The main problem we are working on is to find the most important piece of data that the student needs,” says Serban.
Korbi is a student as well as a teacher. The scheme adapts the treatments it provides to human users over time as it discovers what works. For thousands of people at once, the fact that the instrument can teach at scale often means that it has access to vast volumes of the knowledge it uses to reinforce it.
“We let it be figured out by the AI algorithm from its data,” Serban says. Learning from the learners is much of what it does. Students teach them to do better.