Machine Teaching: Frenemy of Machine Learning
Monday, May 12, 2014 - 4:00pm - 4:50pm
KEC 1001
Xiaojin (Jerry) Zhu
Associate Professor
Department of Computer Sciences
University of Wisconsin-Madison
Abstract:
Consider the inverse problem of machine learning: a teacher knows a learner's learning
algorithm and wants to construct the smallest (non-iid) training set to guide the learner
to a specific target model. This problem, which we call machine teaching, is about
designing the optimal "lesson" to maximally influence the learner. One
application of machine teaching is in the security of machine learning systems that
accept online training data. Here, the teacher is an attacker who can manipulate the
training data. The computational problem is for the attacker to identify the minimum-cost
manipulation so that the machine learner will be misled to a model that is beneficial to
the attacker. A more friendly application of machine teaching is in education, where the
teacher wishes to design the best lesson for a human student. Note that in both
applications the teacher may only interact with the learner via the training data. I will
introduce an optimization-based framework for machi!
ne teaching, balancing the goals of "teaching well" and "minimizing teaching
effort." For certain learners, machine teaching has a closed-form solution. But in general the
optimization problem is combinatorial. I will discuss two approximate solution techniques based on
conjugate duality and submodularity. I will also discuss the relation between machine teaching,
active learning, and teaching dimensions. Finally, I demonstrate the application of machine
teaching with attacks on several popular machine learning models.
Biography:
Xiaojin Zhu is an Associate Professor in the Department of Computer Sciences at the University of Wisconsin-Madison. Dr. Zhu received his B.S. and M.S. degrees in Computer Science from Shanghai Jiao Tong University in 1993 and 1996, respectively, and a 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, and best paper awards at ICML, ECML/PKDD, and SIGSOFT. His research interest is in machine learning, with applications in natural language processing, cognitive science, and social media analysis.
_______________________________________________
Colloquium mailing list
[email protected]
https://secure.engr.oregonstate.edu/mailman/listinfo/colloquium
