Scaling Memory-Augmented Neural Networks with Sparse Reads and Writes