STAT 591: Modern topics in machine learning: Multiway
	classification, preferences, intransitivity

	Marina Meila mmp@stat.washington.edu
	Jeff Bilmes bilmes@ee.washington.edu

	PDL C-301  Wednesday 12:30-1:50

	2 credits

	www.ms.washington.edu/stat591/fall06 (to appear)

	* Supervised learning - classification, ranking and all that
	* Support vector machine classifiers
	    with emphasis on multiclass classification
	* Boosting classifiers
	* A general framework for multiclass and ranking tasks
	* Sources of intransitivity in classification, ranking and
 	  economic/social choice, preference relationships
	* Models of intransitivity. Multi-tiered, Relationships with boosting,
	  permutations, mixture models, and various other 
          unusual suspects

This course is an incursion into supervised learning beyond binary
classification. Multiway classification, aka making decisions with
multiple choices, is a much richer task than two-way classification, and
a harder one too. Many advances in this field, as well as an interest in
related problems like ordering a set of options (instead of choosing
one), have occurred very recently. In this course, we shall investigate
the issues that make multi-way decisions different from binary ones, and
we will survey an array of paradigms for building discriminative
classification and ranking systems.

In doing so, we will introduce the fundamental concepts of machine
learning and building models for data, as well some very successful and
versatile learning methods like boosting and large margin. We will
also discuss another aspect of modern machine learning: the penchant
for non-parametric methods, their advantages, and challenges. 

Non-parametric modeling broadly denotes learning from data beyond
"fitting a function". Rather, learning is done in a framework flexible
enough to let the data dictate the complexity and "shape" of a
model. For instance, a support vector (SV) classifier is defined by a
set of relevant points from the data itself (the support vectors). If a
classification task is hard, the trained SV classifier will have many
vectors, while for an easy problem only a few SV's will be selected. The
number and particular set of SV's are chosen automatically by the
training procedure.

Intransitivity in multiway decision making is gaining
recognition. Essentially, intransitivity means that the outcome of
several related choices may lead to circularities. We will show how
intransitivity can appear in classification as well as in human decision
making (the latter models stemming from the theory of economic and
social choice). We will study from were intransitivity may naturally
arise, and develop methods and models to handle and exploit
intransitivity.

This course assumes some basic understanding of machine learning,
pattern recognition or probability and statistics.