Scaling Memory-Augmented Neural Networks with Sparse Reads and Writes

Sparse Access Memory is like a differentiable hash table. I'm wondering how many new differentiable data structures could be invented in order to augment deep neural nets?