Sunday, Nov 20, 2005, 11:15-12:15
Room 309
--------------------------------------------------------------------------------
John Langford
TTI
Title:
Learning Reductions Formalize Learning Intuition
Abstract:
Learning reductions take a learning problem of type A
transform to a learning problem of
type B and then analyze (and
algorithmically optimize) how
performance e on the created problem
implies performance on problem A. The central claim here is that this
process formalizes the intuitions
of several empirically successful
learning algorithms including "error
correcting output
codes" (Dietterich
and Bakiri), "end-to-end learning" (Yann LeCun), and
constraint-based structured
prediction (Dan Roth). This
formalization
makes it easy to describe the "why"
of these algorithms, to propose
improved algorithms, and to
manufacture new learning algorithms with a
reasonable expectation of empirical
success. I will discuss the
mathematical foundations of
learning reductions and give several
examples of how they work.