What is the state of neural network pruning?

Model size reduction is incredibly important for getting ML running on people's devices. This paper is an awesome summary of the existing research, and its relative lack of coherence. :)