|
I am a second year phd students and I really want to learn graphic models, at the same time, I have some graduate students who also wanting to learn this. I am hoping to host a reading group on graphic topics. Could anyone kind enough to recommend some good resources? |
|
Though by no means definitive, here are a few pointers that should be a good place to start with:
Among books, the new book by Koller & Freidman is nice and comprehensive (a recently offered course that followed this book) |
|
If you have continuous variables with a roughly Gaussian distribution, I have found that Gaussian graphical models give a good framework. In particular, recent work on sparse inverse covariance estimation gives good algorithmic tools. For this theoretical setting, the book "Graphical Models", by S. Lauritzen (1996) is a really good read. It is quite theoretical, but well-written. I realize that the question was for a 'student', so I should add that this is probably not the book to start with, and I would start with scanning through the Jordan book mentioned above.
This answer is marked "community wiki".
|
|
Jordan/Weiss's "Graphical models: Probabilistic inference" gives a short and self-contained introduction to inference in graphical models. Bishop's PRML has an easy to follow derivation of sum product algorithm. Koller/Friedman's latest book gives detailed description of Junction Tree algorithm. Mark Paskin's thesis is another self-contained introduction to Junction Tree algorithm. IMHO, the best way to learn some of the issues in graphical models is to try to derive key quantities yourself. Notation-wise, the easiest setting is an Ising model where xi's are binary and joint probability is of the form p(x)propto exp(x' J x + h x). Then common algorithms like sum product start to look particularly compact, ie, page 3 of Montanari's lecture notes 1
+1 for Bishop's PRML. If you're a student and want to learn about graphical models than go for that text. As much as I think Graphical Models, Exponential Families, and Variational Inference is an amazing piece of work, it is nor a book, nor a course nor a survey of graphical models but rather a very specific piece of research connecting graphical models to optimization.
(Aug 31 '10 at 16:24)
Jurgen
I've found PRML's text on graphical models to be confusing. More than one student after reading it (on my request) said "cool", but had no idea how he could go about implementing one of those things. I think it's missing a clear, small, and concise text on graphical models showing lots of examples, both of the model and of the inference algorithms and clear experiments highlighting their advantagens and disadvantages (like "a guided tour of graphical models and their applications in vision, signal processing, and nlp" or something like that).
(Aug 31 '10 at 21:02)
Alexandre Passos ♦
|
|
If you're interested in graphical models for nlp specifically, you should also look at
The Koller & Friedman book is very comprehensive on the theory behind graphical models per se, but is very light on actual examples. It shows no sampling algorithms (though you can easily derive most from papers, but still), for example, and barely mentions most optimization techniques you need to actually solve inference and estimation problems. M³Ns are also very superficially explained, as are structured SVMs in general. You might enjoy David Mackay's book if you aren't familiar with probability or machine learning, but it is not state of the art and very opinionated. |
|
In the answers to the question Good Freely Available Textbooks on Machine Learning you will find a number of texts that cover graphical models. Of the links given there the most appropriate are, I think, Bayesian Reasoning and Machine Learning is really good,and suitable for newbies. Jordan's 300 pages survey is suitable for people know this field a little bit.
(Aug 24 '10 at 19:56)
Zhibo Xiao
|
