Moving Beyond Labeled Input/Output Pairs in Machine Learning

Dr. Jude Shavlik 
Computer Science Department         
University of Wisconsin - Madison

Friday, February 28, 2003
10:00 a.m.

Life Science 185

ABSTRACT

The "information pipeline" between humans and machine-learning systems
currently is very narrow.  Although there have been attempts to broaden
what we provide to learning systems as part of the training process, the
most-investigated approach by far has been to only provide the learner
with a set of sample input-output pairs of the function we want it to
acquire.  While limited, this narrow perspective has proven successful on
many real-world problems.  However, it seems likely that allowing humans
to provide a wider range of knowledge to machine learners will lead to
computer systems that have a deeper understanding of their training
examples, learn faster and more accurately, are easier for humans to
interact with and understand, and apply to more complex, dynamic tasks.
In this talk I will discuss some of the work done at UW-Madison over the
past decade that allows one to provide additional information to learning
systems.  Our most recent work that allows one to provide background
knowledge to support-vector machines will serve as a detailed case study.
(However, no prior knowledge of support-vector machines by the audience
will be assumed.)