General
General methods are almost always better than imposing human knowledge on the system (e.g. Alpha Go Zero, Image Processing with edges vs deep learning). The only case where human knowledge might help a bit is when you have a product that is not working without a lot of data and that data is only obtainable during operation, for example self driving cars. Then, it might be helpful to develop a lower capability system with the help of human knowledge to have a sellable system and get the data to train the system from scratch without human knowledge (e.g. Tesla FSD)