By "how to train deep networks" are you referring to unsupervised pre-training, or the discovery/development of faster parallel training? (GPUs and distributed ASGD)