Subject: Next Topic: Latent Dirichlet Allocation...
The next topic for the class will be probabilistic graphical models. The tricky part is that it is not clear to me what I can expect as a common denominator background on graphical models among all of you.
I am thinking that instead of trying to do a breadth-sweep, it is better to go depth-first and learn one specific (and as it turns out, highly useful) idea in the class. This turns out to be latent dirichlet allocation--a highly useful topic modeling tool. (those of you who know latent semantic indexing or PCA might think of LDA as a significantly more general topic modeling tool).
The interesting part about LDA is that understanding it requires a jump from 471-level graphical models (in particular, you will use plate models, continuous random variables, learning with incomplete data, bayesian learning, more sophiticated approximate inference techniques etc.).
At the same time, rather than understand these in abstract, you will understand them on an "as needed" basis to understand LDA.
To complement the conceptual understanding of LDA, we will get a project that involves using an existing LDA tool box to analyze some text corpora.
LDA is not described in the R&N textbook. We will instead read the original LDA paper:
as well as an online video tutorial on LDA
While all of this does make life challenging I think it is going to be more representative of a true graduate class (as in "not even the professor completely understands the topic"..)
If you have any concerns, let me know asap.
Rao