|
|
|
CSci 8980: Topics in Machine Learning, Spring 2006
Instructor:
Arindam Banerjee
Time and Place: T Th
12:45-2:00pm, Room 28, Peik Hall (Note Room Change)
Office Hours: T Th
2:15-3:15pm, EE/CS 6-213, or by appointment
Email: my_last_name at cs.umn.edu, add [8980]
to subject line
Overview:
Have you ever been through a situation where you had
to choose one out of multiple options, and you later wished you had
chosen something else? Such situations can arise in anything from
choosing a bad slot-machine in a casino, to putting your money on
the wrong stock, to not registering for the right course. At the end
of the day (or semester), you regret your choice and hope to learn
from your mistakes.
In this course, we will learn how to effectively learn from mistakes
so that eventually we have no regret. In particular, we will study
learning in repeated games, and its connections to popular machine
learning techniques such as boosting and large-margin methods. On the
computational game-theory side, we will study no-regret algorithms for
achieving equilibrium in multi-party as well as graphical games. On
the machine learning side, we will develop an understanding of how
boosting and ensemble methods are intimately connected to learning in
games. We will also delve deeper into large-margin methods and convex
optimization that play a central role in modern predictive modeling,
including supervised and unsupervised learning. Applications of these
ideas to data mining, portfolio design, structured prediction, social
networks etc., will be studied.
Topics:
Online Learning, Computational Games, Boosting, Convex Optimization,
Large Margin Methods, Learning Theory, Mixture Models, Social Networks.
Text:
No textbook. The course would be based on technical papers and handouts.
Course Work:
The course will be based on paper readings, class
presentations and discussions. Students will be required to submit a
brief review of the papers in the reading list. In addition, there
will be a course project that can be done individually or in groups
(at most 3 students). In case of a group project, contributions of
every student must be clearly stated.
Grading:
Project: 45%, Paper Reviews: 30%, Presentation: 15%, Participation: 10%
Prerequisites:
Graduate standing; basic background in pattern
recognition and/or data mining (equivalent of CSci 5521 and CSci
5523), or consent of instructor. Knowledge of statistics or game
theory is a plus.
|
|
