Yes [1].

[1] https://arxiv.org/abs/1801.04016

For me Bayesian networks are a tool of thought and I think it's worthy to learn them in the same way it's worthy to learn to sketch functions with pen and paper.

Checkout http://edwardlib.org/ and https://www.tensorflow.org/api_docs/python/tf/contrib/bayesf....

Also Bayesian Methods for Hackers: https://news.ycombinator.com/item?id=16482421

Yes, I think so. PGM or some improvement on that is relevant to doing things like reasoning correctly based on evidence.

I envision more sophisticated AGI systems to use DNN or other NN techniques to learn about the world, and be able to take in uncertain input and make sense of it. PGM or similar would then be used to correctly (in the mathematical sense) reason about what to do to accomplish the agent's goals.

Also graphical models are combined with deep learning techniques for SOTA performance in area like Named Entity Recognition.