Work done at Google, in collaboration with Eyvind Niklasson and Alexander Mordvintsev.

Meta learning a learning protocol that can be used instead of SGD, Adam, etc, to train feed-forward neural networks. At least, in theory.

See the Arxiv preprint and the code base. The code base contains colabs that show how to use it!

My first first author work. Very fundamental in nature and small in scope. Still, it is 9 pages long.