Nice paper from Google/DeepMind building on the Lottery Ticket Hypothesis paper from earlier this year with a new approach to a train a sparse network from scratch without having to train the network then prune its connections.
https://arxiv.org/pdf/1911.11134v1.pdf
Code