What can the working (pure) mathematician expect from deep learning?