There are at least two huge advantages of the decomposition by function approach compared to the black box solution space search approach: - The former can be

There are at least two huge advantages of the decomposition by function approach compared to the black box solution space search approach: - The former can be debugged. If your image analysis system embarrassingly tags black people as gorillas, you can go in and find the bug and fix it. That's not so easy if you're using a black box model. - The black box approach is tremendously enervating. You code up a neural network architecture, launch a bunch of cloud GPU instances, and start training. If the results are bad... you try a new architecture. In the function decomposition world, you can actually use your knowledge and understanding of the system as an engineer to figure out what went wrong and why.