The Gong Show

Probabilistic Graphical Models

I started working through the video lectures in Daphne Koller’s Probabilistic Graphical Models (PGM) class yesterday.  I wrote a bunch last fall about how much I enjoyed Andrew Ng’s Machine Learning (ML) online class.  PGM is not technically a sequel to the ML class, but there are a number of overlapping concepts, and all the homework assignments are to be completed in Octave, which is the same as the ML class.  Many students at Stanford take the PGM class after taking the ML class, so it’s a successor in practice.

Whether you know it or not, you use PGMs everyday.  The most common example is Google.  The Google search experience was originally powered by Pagerank (it’s now powered by hundreds of complex heuristics beyond Pagerank), which is a specific type of PGM called a Markov Chain. Pagerank determines the probability that you’ll end up on a given page when surfing the web, which is then used as a proxy for that page’s authority.

Another example of real world PGMs you use everyday is the recommended friends list in Facebook or LinkedIn.  Every wonder how Facebook knows to recommend specific people that are eerily accurate? That’s a PGM at work where the graph used for modeling is the social graph, and the probabilities are the likelihood that an edge should be formed in a place where one does not currently exist.

Professor Koller teaches at a quick clip, a bit faster than Professor Ng, and she assume a bit more background knowledge of her audience, which so far is a net-positive experience. The courseware (provided by Coursera) that powers this massive open online course (MOOC — yes, it’s become an acronym) has been given an upgrade since Professor Ng taught the ML class, and it’s all for the better (except, oddly, the search function is missing from the forum page now… am I just missing it?).

I’m blogging about the PGM course for two reasons. 

1. Come take this class with me.  It’s amazing.  I’m eager for more content to be released.  Taking ML class last fall was an incredibly rewarding experience, and I’m sure PGM will live up to my high expectations.
2. By writing about this, it’s a way of publicly holding myself accountable to finish the course.  It will likely be 10 hours of work per week for 10-12 weeks, and that’s not a trivial commitment.  I want to make sure I finish it, and the best way to do that is to tell everyone I know that I’m taking the class, so I can’t back out later when I inevitably become too busy.

A big thanks to Professor Koller for taking the time to build this course.  She is the same person that wrote this awesome NYT Essay about MOOCs and their disruptive role in the future of education.