We understand machine learning itself, it's the problem space that needs to be understood. That's why it is such a black box - you never know the complexity of the model you need to build before you tackle it.