[ponderingHplus]

I attended my first NIPS this year, and found Juergen to be a very engaging speaker, with the RNN symposium organized by him and his colleagues being my favorite part of the conference. A popular phrase that was being thrown around during the conference was "learning to learn" or "meta learning", with one of the papers even being titled "learning to learn by gradient descent by gradient descent"[1] Juergen seemed very passionate about the subject and he gave a cool talk around his Godel Machine[2], and sparked interesting conversation during the panel discussion. I wouldn't be surprised if "learning to learn" or "meta learning" replaces "deep learning" as the AI-word of 2017.

[1]https://papers.nips.cc/paper/6461-learning-to-learn-by-gradi...

[2]https://en.wikipedia.org/wiki/G%C3%B6del_machine

[deepnotderp]

Learning to learn is a deepmind paper, not schmidhuber.

[ponderingHplus]

Agreed, but the linked paper was one of the more talked about ones during the conference, and has a fairly accessible discussion on the topic, including a history of related work in section 1.2 with references to many of Jurgen's papers.