Learning From Data is the course I recommend to anyone who can already train a model but has a nagging sense they do not really know why it generalizes. Yaser Abu-Mostafa spends the term on the deep question underneath all of machine learning, which is how a finite amount of data can ever justify a claim about data you have not seen, and he answers it properly with the theory of generalization, the VC dimension, the bias-variance tradeoff, regularization, and validation. What makes it special is the teaching. Abu-Mostafa builds intuition first and reaches for the formalism only once you already feel why it has to be true, so results that look forbidding in a textbook arrive feeling almost obvious, and the famous lecture on overfitting alone is worth the price of admission, which is nothing.
It is rigorous without being cold, and the problems are where the understanding actually settles in, so treat it as a course to work through rather than a series to watch. The caveats are honest. This is mathematically demanding and assumes real comfort with probability and linear algebra, so it is not a gentle introduction and it will frustrate anyone hoping to skip the equations. It also predates the deep learning era, so you will not find transformers or modern architectures here, and that is fine because the principles it teaches apply to them anyway, but you should not expect coverage of today's models.
And it is theory first, so you finish it understanding learning deeply rather than holding a fresh pile of trained projects. Taken for what it is, the theoretical backbone that almost every practical course quietly leaves out, it is close to perfect, and I would tell any serious learner to take a hands-on course first and then come here to find out why all of it works.