The Unreasonable Success of Intelligent Tutoring Systems

Friday, January 5, 2018: 3:30 PM
Roosevelt Room 4 (Marriott Wardman Park)
Shreeharsh Kelkar, University of California, Berkeley
Intelligent Tutoring Systems (ITSes) are a particular class of Artificial Intelligence (AI)

systems, dating back to the 1960s, which aim to simulate a human tutor; the system

presents information to the student and offers examples for practice. As the students

practice, the system computes their knowledge state from their actions, and offers

dynamically generated hints and scaffolds, that take into account both the current

knowledge state of the student and the knowledge/skills that the student is supposed to

learn. For a project that did not make much headway in the 1960s, the ITS community,

with its center of gravity at Carnegie Mellon University (CMU), has been startlingly

successful, culminating with the establishment in 1998 of the for-profit spinoff Carnegie

Learning—whose signature product, Cognitive Tutor, is now used in more than 1000

schools across the US. In 2005, the US National Science Foundation (NSF) awarded

Carnegie Learning, CMU and the University of Pittsburgh a 50 million grant (over 10

years) to form the Pittsburgh Science of Learning Center (PSLC). The PSLC, in turn,

has been the key institution that sustains at least three research communities (in terms

of sponsorship, data, and people): Artificial Intelligence in Education (AIED), the

learning sciences (LS) and Education Data Mining (EDM) communities. ITSes are a

perfect bridge across the multiple social contexts of Artificial Intelligence: they have links

to cognitive science, expert systems (i.e. "good old-fashioned artificial intelligence" or

GOFAI), and machine learning; they span both university research labs and for-profit

corporations, they are also experimental systems built to produce knowledge about

learning as well as practical systems meant to help students learn concepts.

ITS systems then represent an unusual success of the much maligned GOFAI; yet, a

closer examination reveals that their makers took several unconventional steps to make

them successful: in order to get Pittsburgh schools to use these systems, they had to

devolve substantial authority to teachers, necessitating a fundamental reconception of

the ITS itself. Rather than seeing the tutoring system as mimicking a human tutor—a

conceptualization that is classically symbolic AI in its framing—they came to realize that

it was best to frame it instead as a learning aids for both teachers and students working

together. Along the way, however, the makers of ITSes were able to secure

professional authority as “learning scientists” whose expertise on learning and cognitive

science was separate and more important than expertise in subject-matter or

technology.

Previous Presentation | Next Presentation >>