Machine teaching unites computer science and psychology to illuminate how people learn

Machine-teaching researchers discuss their project

Human learning is a complex, sometimes mysterious process.  Most of us have had experiences where we have struggled to learn something new, but also times when we’ve picked something up almost effortlessly.

What if a fusion of computer science and psychology could help us understand more about how people learn, making it possible to design ideal lessons?

That long-range goal is moving towards reality thanks to an effort led by professors in the University of Wisconsin-Madison departments of computer science, psychology and educational psychology.  Their collaborative research aims to break new ground in what computer scientist Jerry Zhu calls “machine teaching”—a twist on the more familiar concept of machine learning.

“My hope is that machine teaching has an impact on the educational world.  It’s quite different from how people usually think about education,” says Zhu (pictured at right). “It will give us optimal, personalized lessons for real, human students.” Jerry Zhu

Machine learning is a well-established subfield of computer science in which experts develop mathematical tools to help computers learn from data and detect patterns.  The machine learner (the computer) is like a student.  The goal of machine learning is to develop models that will prove useful in the future when dealing with large, often unwieldly data sets.  Practical tasks like speech recognition are aided by machine learning.

Machine teaching--a term coined by Zhu--turns this concept on its ear.  Rather than dealing with pools of data and not knowing at the outset what patterns might be revealed through analysis, the researcher in a machine teaching arrangement already knows what knowledge or models he or she wants to impress upon the learner.  The goal is to reach this end using the smallest set of training data possible.

While this work is still in its early stages, its potential impact upon education is immense.  It also opens up many theoretical questions (for example, what is the absolute smallest training set of data you can construct to make learning happen?).  

Zhu presented some of this work earlier this year in Austin, Texas, at the 29th annual Conference on Artificial Intelligence, organized by the Association for the Advancement of Artificial Intelligence.

Zhu’s conference paper won the Blue Sky Ideas Prize.  Sponsored by the Computing Community Consortium (CCC), the Blue Sky initiative is designed to “seek out papers that present ideas and visions that can stimulate the research community to pursue new directions,” according to the CCC.

Timothy T. Rogers, a professor of cognitive psychology at UW-Madison and one of Zhu’s collaborators, explains how computer science and psychology come together.

Timothy T. Rogers“In order for the machine-teaching approach to work, it needs a good model of how the learner behaves—that is, how the learner’s behavior changes with different kinds of learning or practice experiences,” says Rogers (pictured at left).  “Also, the model needs to be computational; it has to be able to make concrete, quantitative predictions about the learner’s behavior.”

"Ultimately, we hope that the work can be used to help teachers develop lesson plans and curricula that promote learning in a wide variety of fields," Rogers says, citing math, science and reading as examples.  “And, just as important, the effort to bring cognitive models of learning to bear on real-world problems is bound to lead to important new advances in our understanding of how people learn generally."

The benefits of machine teaching extended beyond education.  The emerging field also has potential for improved cybersecurity.

Machine-teaching principles can help computer scientists ward off attacks known as “data poisoning,” in which an attacker influences data coming into a system with the intent of corrupting it.

Although data poisoning may not be a widespread phenomenon at present, any system that takes input and adapts based on that input—like a spam filter—is potentially vulnerable.

A two-year seed grant from the University of Wisconsin-Madison Graduate School currently supports this work.  Future funding from outside sources will be sought.

"With machine teaching, it’s conceptually easy, but quite challenging to implement in the real world.  It’s a major undertaking," says Zhu.

In addition to Zhu and Rogers, the UW-Madison research team includes computer science professors Michael Ferris, Bilge Mutlu and Stephen Wright; engineering professor Rob Nowak; psychology professor Martha Alibali; and educational psychology professors Percival Matthews and Martina Rau.

Computer science graduate students Gorune Ohannessian and Ayon Sen are also involved, as is recent graduate Shike Mei.  Outside of Madison, collaborators include Bradley Love, a professor of experimental psychology at University College London, and Ji Liu, a computer science assistant professor at the University of Rochester in New York.  Liu earned his Ph.D. at UW-Madison last year.

Machine teaching probes fundamental mathematical and scientific concepts.  In part because of that, the team’s research is open-ended at this stage.  To follow the team’s progress and learn more, visit Zhu’s website.

[Pictured at top: Collaborators discuss their work together. From left to right: Martha Alibali, Percival Matthews, Jerry Zhu and Tim Rogers.]