Distributed training of very large neural networks