UTCS Colloquium/AI: Jesse Davis/University of Washington Statistical Relational Learning Diagnosing Breast Cancer and Transfer Learning TAY 3.128 Friday September 12 2008 11:00 a.m.

Contact Name: 
Jenna Whitney
Date: 
Sep 12, 2008 11:00am - 12:00pm


There is a signup schedule for this event (UT EID required).

Type of Talk: UTCS Colloquium/AI

Speaker/Affiliation: Jesse Davis

/University of Washington

Date/Time: Friday September 12 2008 11

:00 a.m.

Location: TAY 3.128

Host: Raymond Mooney

Ta

lk Title: Statistical Relational Learning
Diagnosing Breast Cancer an

d Transfer Learning

Talk Abstract:
Standard inductive learning ma

kes two key assumptions
about the structure of the data. First it requ

ires that all
examples are independent and identically distributed (iid

).
Second it requires that the training and test instances come
fr

om the same distribution. Decades of research have
produced many sophis

ticated techniques for solving this
task. Unfortunately in real applic

ations these assumptions
are often violated. In the first part of this

talk I will motivate
the need to handle non-iid data through the concr

ete task of
predicting whether an abnormality on a mammogram is
mal

ignant. I will describe the SAYU algorithm which
automatically constru

cts relational features. Our system
makes significantly more accurate p

redictions than both
radiologists and other machine learning techniques
on this
task. Furthermore we identified a novel feature that is
i

ndicative of malignancy. In the second part of this talk
I will discus

s a transfer learning algorithm that removes
both restrictions made by

standard inductive learners. In
shallow transfer test instances are fr

om the same domain
but have a different distribution. In deep transfer
test
instances are from a different domain entirely (i.e.
descri

bed by different predicates). Humans routinely
perform deep transfer b

ut few learning systems are
capable of it. I will describe an approach

based on a
form of second-order Markov logic which discovers
struc

tural regularities in the source domain in the form
of Markov logic for

mulas with predicate variables and
instantiates these formulas with pr

edicates from the target
domain. Using this approach we have successfu

lly transferred
learned knowledge between a molecular biology domain an

d
a Web one. The discovered patterns include broadly useful
properti

es of predicates like symmetry and transitivity and
relations among p

redicates like various forms of homophily.

Speaker Bio:
Jesse Da

vis is a post-doctoral researcher at the University
of Washington. He r

eceived his Ph.D in computer science at
the University of Wisconsin - M

adison in 2007 and a B.A. in
computer science from Williams College in

2002. His research
interests include statistical relational learning t

ransfer learning
inductive logic programming and data mining for biome

dical domains.