I just finished the three-part series of Probabilistic Graphical Models courses from Stanford over on Coursera. I got in to them because I realized there is a certain class of problem for which the standard supervised learning approaches don't apply, for which graph search algorithms don't work, problems that don't look like RL control problems, that don't even exactly look like the kind of clustering I came to call "unsupervised learning".
In my AI courses in the Institute, we talked briefly about Bayes Nets, but it was almost as if professors considered that preamble to hotter topics like Neural Nets. Meanwhile I heard about "Expectation Maximization" and "Inference" and "Maximum Likelihood Estimation" all the time, like I was supposed to know what they were talking about. It frustrated me not to be able to remember statistics well enough to feel these things, so I decided to fill the hole by delving deeper in to PGMs.
Throughout, Koller gives examples of how to apply PGMs to things like image segmentation and speech recognition, examples that seem completely dated now because we have CNNs and LSTMs, even deep nets that encode notions of uncertainty about their beliefs.
I gather PGMs are good when:
- You know the structure of the problem and can encode domain knowledge that way.
- You need a generative model.
- You want to learn more than just one X -> Y mapping, when you instead need a more general-purpose model that can be queried from several sides to answer different kinds of questions.
- You want to feed the model inputs that look more like probability distributions than like samples.
What else are they good for? Have they not been outstripped by more advanced methods for lots of problems now? In which domains or for which specific kinds of problem are they still king? How are they complementary to modern advanced methods? Should I dedicate time to reading any of Koller & Friedman's massive tome on this subject? How dated is this set of MOOCs?