At the end of every year, I like to take a look back at the different trends or papers that inspired me the most. As a researcher in the field, I find it can be quite productive to take a deeper look at where I think the research community has made surprising progress or to identify areas where, perhaps unexpectedly, we did not advance. I’ve put together a post in which I give my perspective on the state of the field from this past year. The post is no doubt a biased sample of what I think is progress, but I hope it stimulates discussion about which subfields evolved or what priorities unexpectedly shifted in 2019.

Link to post: http://www.cachestocaches.com/2019/12/my-state-of-the-field/

A short summary/outline of my post for discussion:

• From AlphaZero to MuZero MuZero picks up where AlphaZero left off a couple years ago and makes significant advances using a learned model to enable rollouts without planning in pixel space.
• Representation Learning I’m particularly excited to see how recent progress in representation learning (like advances in “entity abstraction”) will help to blur the lines between black-box deep learning and old-school symbolic AI & classical planning.
• Supervised Computer Vision Research Cools (slightly) Research in this space has slowed, but related techniques, like network pruning and network compilation, have taken off this past year.
• Maturing Technologies
  • Graph Neural Networks
  • Explainable & Interpretable AI
  • Simulation Tools & Sim-to-Real
• Bittersweet Lessons No discussion of 2019 would be complete without a conversation about Rich Sutton’s “The Bitter Lesson” post and rebuttals.

What do you think were the most interesting advancements or shifts this year?