By Joshua Preston | Sep. 15, 2019 |
Georgia Tech seeded a national change in higher education in bringing its Master of Science in Computer Science degree program to the web at a fraction of the cost as the same on-campus degree.
The online program also introduced an artificially intelligent teaching assistant (virtual TA) that is now used in several courses.
The AI agent Jill Watson – introduced in 2016 to answer common course questions – is marking its fourth year with several milestones that have the potential to lead to transformative global changes in education.
Researchers started with computing courses in introducing the virtual TA, which can correctly answer course questions from students whenever they need help. Now, the first science course – Introduction to Biology – is using Jill Watson to answer student questions.
“This is a significant test for our AI system,” said Varsha Achar, MSCS alumna and a member of the Jill Watson team. “Jill has interacted with multiple thousands of students in online computing courses and is now tackling a very different subject that will show how the system can be used much more broadly.”
Not only is Jill answering questions related to science material, but the agent’s first day in a residential classroom was this semester (biology and CS courses on the Georgia Tech campus). It’s also being deployed in a record five courses as a TA. Two was the previous record for a single semester.
The research team is confident Jill is ready. Members of the Design & Intelligence Lab, directed by School of Interactive Computing Professor Ashok Goel, have spent countless hours tweaking Jill’s code and the best approach to scale the TA to any college course.
“The success of Jill Watson lies in answering questions in an almost human-like manner and how it has come to do so,” said Goel.
Jill’s very first online students three years ago didn’t know an AI was answering a portion of their questions, and many were surprised when they were told at the end of the online course. Building a believable agent – years in the making – has required many different approaches and data for training the system.
The growing course load for the virtual TA is possible, in part, through a document that students get on the very first day of class – the course syllabus.
Achar, the chief architect of the knowledge bases Jill uses to answer questions, has created these sources based on data in the syllabi. This includes structured data (e.g. deadlines, test dates, submission processes) and unstructured data (e.g. disability services, policies on late assignments).
It’s a method that is proving key to Jill’s growth.
“We don’t include anything for Jill to answer that’s not in the syllabus, except information on relevant institute policies,” said Achar. “That’s important to keeping the integrity of the system. I could easily google extra information that might be needed, but we’re not the experts on the classes.”
A New Vision for AIs in the Classroom
In a year of firsts for Jill, the most critical is a new approach to building the agent, one that might prove to accelerate the AI race in higher education.
The biology course is using not one, but two Jills – the first answers common questions and another version of Jill partners with a powerful AI agent that is an expert in ecology.
This second AI agent, VERA , or virtual ecological research assistant, was developed by members of the same lab and is used to introduce students to experimentation with ecosystems and food chains.
To get undergrads quickly up-to-speed on how to use VERA’s modeling and simulation capabilities, students are able ask Jill questions on how to use VERA. If students get stuck in the scientific discovery process, they can ask Jill how to, for example, put a predator into the VERA system and change population size, rather than thumbing through a technical manual.
This new version of Jill was created using the VERA tutorial, which students may never need to crack now. It proves that the TA system can adapt from a syllabus to a tutorial, something that opens up the possibilities for creating even more complex agents, said Achar.
Right now, it takes Achar and the team a little over 24 hours (including deployment time) to build Jill for a course. They have their sights set on bringing that down to less than a business day and, in time, allowing instructors to build their own versions of Jill.
And using Jill and VERA together hints at the ambitious goals the research team has in leveraging the power of multiple AIs to be able to adapt any course material.
“It’s extremely exciting to be a part of this team,” Achar says. “This is an opportunity to create groundbreaking technology and do things in education that haven’t been done before.”
“We’re able to shape minds and inspire people in countless ways. This work could inspire others to build future AI systems or learn in unprecedented ways. It’s something that I see helping deliver on the promise of lifetime learning.