Parallel Training of Deep Networks with Local Updates