I think a good way to structure a course would be to cover several problems from end-to-end: motivation (what you want to achieve), theory, data preprocessing,

I think a good way to structure a course would be to cover several problems from end-to-end: motivation (what you want to achieve), theory, data preprocessing, algorithm development, and finally setting up a "production grade" system that solves the problem. In my experience learning ML, learning concepts in theory is good and all, but I never really understood the details until I had to implementing the algorithm.