One question that's not explicitly asked on this list: How much preparation does the training data need?

Knowledge representation is still one of the main bottlenecks on AI/ML in general, a deep learning network is not going to prepare its food by itself.