My pleb take is modern machine learning is just glorified complexity theory. Really all you are doing is solving hard problems in learning models by designing a

My pleb take is modern machine learning is just glorified complexity theory. Really all you are doing is solving hard problems in learning models by designing a continuous process, deterministic or randomized, showing it has desired properties of arriving at some optimal solution or probability distribution and then deriving a discrete algorithm that runs in polynomial-time because math optimization problems in general are NP-hard. Now we have non-convex neural network models which require non-convex optimization which to me (again a pleb take) is just tricks of the trade from complexity theorists adapting the principles of convex optimization to things like gradient descent in deep learning by observing continuous local smoothness of the training objective at the stability edge thus some convex optimization can be used. Why it's not taught instead of the confusing intro courses I'm sure have their reasons but it's another example of following what the universities teach in undergrad is not always the best road map for self-learners.