I had been planning to implement the Differentiable Neural Computer with libdeep this year, but decided against doing that since I didn't want to make the library more complex than it needs to be. But maybe I'll do it as a separate project in 2018. It would also be worth combining this with the capsule network idea for better classification matching on memory locations.

DNC is a bit more interesting than the usual deep learning stuff in that it combines the analog pattern matching of a neural net with the digital processing of a von neumann architecture. This kind of approach is more brain-like and probably more amenable to transfer and multi-domain learning.

https://www.youtube.com/watch?v=6SmN8dt8WeQ