The intention of this repo is to test DNC on several tasks of various sorts to determine if the added memory/context of a DNC will be better for those tasks than the current state of the art.
I got the IMDb and MNIST data sets and ran them through the DNC... Unfortunately, it appears that the DNC only really works for some problems and for others, loss just caps and fails. MNIST hit near state of the art performance though, so there is some potential here for some cool things in the future.
However, I believe there should be some work done to fix a few of the potential bugs of the DNC architecture - which I might be implementing in another repo.
Python 3.7
PyTorch
Jupyter