I think a lot of GOFAI approaches ought to be revisited to see whether they benefit from the new perceptual and decision capabilities of Deep Learning systems. Alex Graves's papers are particularly good at this.

Things like this reinforcement learner for theorem proving are pretty exciting possibilities. https://arxiv.org/pdf/1805.07563v1.pdf

I can recommend his new book, "The book of Why" very highly. Even though I am very familiar with Bayes nets, I discovered that that a lot of progress has been made in that side of AI.