I think it's worth including inference on the list of things that make machine learning difficult. The more complicated your model is, the more computationally difficult it will be to do inference in it, meaning that researchers often have to limit themselves to a much simpler model than they'd actually prefer to use, in order to make inference actually tractable.