Monday Oct 18, WeH 4601, 3:00

Rich Caruana presenting,

TITLE:

Was Billy Ockham Right: Do Smaller Theories Generalize Better?

ABSTRACT:

Ockham's Razor has been a dictum in science for centuries.  Its
formalization in ML theory suggests that the smallest hypothesis
consistent with the data is most likely to be correct.  In this talk
I'll present recent empirical results by Murphy and Pazzani on
decision trees that suggests this is not always true.  They generated
all decision trees consistent with a training set and found that
generalization performance was not a monotonically decreasing function
of complexity.  More specifically, they discovered that in some
domains: 1) the smallest decision trees had worse generalization
performance than those somewhat larger, and 2) generalization
performance sometimes begins to improve as the trees get very large.

-Rich Caruana.

Copies of the paper are available outside of Jean Harpley's office.