Another point to note is that production code needs to meet a higher bar in terms of testing and auditability. It’s one thing to train a NN on a data scientist’s machine and show that on test data, the new system performs 10% better; and another to put what is essentially a black box in production and use that black box to make financial or business decisions which can have far-reaching consequences on data it’s never seen and which is potentially outside of the scope of what it was trained on.