Can Machine Learning Rationalize Simple Human Teaching Behaviors?

Speaker
Xiaojin (Jerry) Zhu
Associate Professor of Computer Science, University of Wisconsin-Madison

When
-

Where
Gates and Hillman Centers 6115

Description

Human teaching is complex. To get a theoretical grip on it, let’s “dumb it down” to a level where current machine learning theory can handle. Consider a concept class of 1D threshold classifiers. To learn a target concept to a given accuracy from i.i.d. training items, a learner needs a polynomial number of items (i.e., a lot). If the learner can do active learning, it needs a logarithmic number of items (i.e., a few)—Note the oracle waits for the learner to pick query items. However, if the oracle picks the training items for the learner, two items suffice: one positive and one negative right around the threshold. Such optimal teaching strategy is captured by a branch of computational learning theory known as teaching dimensions. I will first present some behavioral experiments where humans play the role of the oracle teacher in this setting. Apparently, their teaching behaviors differ dramatically from the optimal strategy predicted by teaching dimensions. In particular, a significant number of human teachers seem to follow the so-called “curriculum learning” strategy by first teaching extreme items at both ends of the 1D range, and then zigzagging toward the threshold. I then present a possible theoretical explanation which suggests that such teaching behaviors might be optimal after all. The explanation is based on risk minimization under certain assumptions. The overall message is that machine learning can offer interesting hypotheses to the study of human teaching and learning, and vice versa.

Speaker's Bio

Xiaojin Zhu is an Associate Professor of Computer Science at the University of Wisconsin-Madison, with affiliate appointments in ECE and Psychology. He received the B.S. and M.S. degrees in Computer Science from Shanghai Jiao Tong University in 1993 and 1996, respectively, and the Ph.D. degree in Language Technologies from Carnegie Mellon University in 2005. He was a research staff member at IBM China Research Laboratory from 1996 to 1998. Dr. Zhu received the National Science Foundation CAREER Award in 2010. His research interest is in statistical machine learning, with applications in natural language processing, cognitive science, and social media data analysis.

Speaker's Website
http://pages.cs.wisc.edu/~jerryzhu/

Host
Burr Settles