EE595A - Dynamic Graphical Models - Winter Quarter 2010

Last updated: $Id: index.html,v 1.18 2010/03/11 06:40:20 bilmes Exp $

Instructor:

Prof. Jeff A. Bilmes --- Email me
Office: 418 EE/CS Bldg., +1 206 221 5236
Office hours: TBD, EEB-418

Announcements

3/10/10: Remember that we have a makeup lecture Thursday, 3/11/2010, from 4-6pm in room EEB-303.
2/10/10: the rest of the lectures slides are not yet here on the public web page, but you can get them in the location I mentioned in class. Email me if you don't know where they are.
2/5/10 Lecture 8 (again mostly whiteboard lecture).
2/3/10 Lecture 7 is available but lecture was mostly white board.
2/2/10 k_best.pdf which describes lecture 6 material is available in the reading area .
2/2/10 Lecture 6 is available.
2/2/10 Lecture 5 is available.
1/12/10 Lecture 4 is available.
1/19/10 Lecture 3 is available.
1/18/10 Lecture 2 is available.
1/15/10 Lecture 1 is available.

Information

Course Format: Four hours of lecture per week (WF 1:30-3:20 LOW-101) per week. Note that there was a time change for this course (it used to be 11:30-1:30pm).

Description: This course will serve as a thorough overview of dynamic graphical models including hidden Markov models, dynamic Bayesian networks, sequential conditional random fields, Kalman filters, switching Kalman filters, and linear and non-linear dynamical systems. We will also discuss various other time-series methods.

For each of the above model types, we will discuss the issues behind and methods for computing various forms of exact and approximate inference, and how this is distinct from the static graphical model case. We will discuss various approximation schemes that are particularly suited to dynamic models. This includes forward/backwards algorithms, temporal junction trees and dynamic triangulation, conditioning and factorization approaches, variational approaches, sampling and particle filter approaches (including importance sampling and Rao-Blackwellization), beam pruning strategies, and multi-pass course-to-fine strategies. We'll discuss the island algorithm and other time-space trade-off strategies.

We'll also cover modeling choices, flexibility and similarity behind each of the models, and their implications.

We will fully cover various forms of learning algorithms for these models, including generative and discriminative training, where the latter includes methods such as MMIE, MDI, and MCE, and the various model-specific approximations (mostly from the speech community), and structured max margin approaches (mostly from the machine learning community). We will compare how they relate to each other, and we'll discuss how the choice of model limits the choice of loss function and corresponding objective function.

We will discuss many applications in speech recognition, natural language processing, bio-informatics, econometrics, and robotics (i.e., localization and mapping), and how the various strategies used in each of these communities relate to each other.

Note that Fall 2009 EE512 was a course that concentrated on static graphical models, while in Winter 2010, we are offering this new course (EE595A) n dynamic graphical models. You may find it useful to gain background by referring back to the material on static graphical models offered during EE512.

Prerequisites: EE512, basic probability, statistics, and random processes (e.g., EE505 or a Stat 5xx class or consent of the instructor). Knowledge of matlab. The course is open to students in all UW departments. If you are in doubt about taking this course, please contact me.

Texts: We will mainly use written material that will be made available on the web page, as well as other printed and online material. Please take a look at the web page for more information.

Grades and Assignments: Grades will be based on a combination of a final project, class attendance, and class participation. There will be no assignewd homeworks, but there will be optional assigned problems that you can work on if you wish.

Final project: The final project will consist of a 4-page paper (conference style) and a 10 minute final project presentation. The project must involve using or dealing mathematically with dynamic graphical models in some way or another. For their application to data, there are a number of software tools (such as BNT or GMTK) which can be used for this purpose, or you may wish to develop your own system. Projects must be done individually.

Homework

There will be no assigned homeworks for the quarter, but there will be some optional problmes assigned occasionally you may work on if you wish. See the lecture notes.

Lecture Slides

Lecture slides will be made available as they are being prepared --- they will probably appear here right before a given lecture, and they will be in PDF format (original source is latex).

Lecture 1 slides
Lecture 2 slides
Lecture 3 slides
Lecture 4 slides
Lecture 5 slides
Lecture 6 slides
Lecture 7 slides
Lecture 8 slides
2/10/10: the other lectures are not yet here on the public web page, but you can get them in the location I mentioned in class. Email me if you don't know where they are.

Readings

The link to some of the readings will be mentioned in class, and some of the other readings will be linked to directly below.

All reading assignments are mentioned in the lecture notes, so see there.

Discussion Board

You can post questions, discussion topics, or general information at this link.

Books and Links

There are many books available that discuss the material that we are covering in this course. A few of the ones that I use regularly, and a few other web links that I think are useful, are included below.

"An Introduction to Bayesian Networks", F.V. Jensen, 1996. A good general introduction to Bayesian networks (out of print, but available in the library).
"Bayesian Networks and Decision Graphs", F.V. Jensen, 2001. Another good general introduction to Bayesian networks (not out of print).
"Graphical Models", S.L. Lauritzen. Oxford, 1996. A very complete, theoretically precise, but dense text, authored by one of the field's leading authorities.
"Probabilistic Networks and Expert Systems", R.G. Cowell, A.P. Dawid, S.L. Lauritzen, and D.J. Spiegelhalter. 1999. Similar to the previous text, but includes more material on inference, applications, and other general problems.
"Artificial Intelligence: A Modern Approach: 2nd Edition", S. Russel and P. Norvig, 2003. Has a nice introductory chapter on Bayesian networks.
"Learning in Graphical Models", Ed. by M.I. Jordan. An excellent collection of recent research papers compiled by Mike Jordan, one of the leading experts in this field.
"Probabilistic Reasoning in Intelligent Systems", J. Pearl. 1998. A classic early text by one of the founders of the field. Pearl is credited with inventing the term "Bayesian networks".
"Causality", J. Pearl. 2000. A relatively newer text by Pearl specifically on causality, and causal modeling. A second edition of this book is out in 2009.
"Pattern Classification", R. Duda, P. Hart and D. Stork (the text used for 596I). The original text (published in 1973) is still widely read.
"The Elements of Statistical Learning: Data Mining, Inference, and Prediction", Hastie, Tibshirani, and Friedman. There is a 2nd edition that came out in 2008.
"Neural Networks for Pattern Recognition", by C. Bishop, 1996. (available now in the UW bookstore). This book mainly contains background material, but has become a classic text in pattern recognition even though "neural networks" is in the title, and is worth reading if you plan to do any work at all in pattern recognition.
Another book by C. Bishop that came out in 2006 is "Pattern Recognition and Machine Learning", which is here. This book also contains a nice overview chapter on graphical models and Bayesian networks.