Some teaching survives its own obsolescence, and this is a good example. Kilian Weinberger recorded his Cornell machine learning course in 2018 and put it on YouTube with the lecture notes on the course site, and years later it is still the series I point people towards when they want to actually understand classical machine learning rather than operate it. The reason is the lecturing. Weinberger does something that sounds simple and is remarkably rare, which is that he explains why an algorithm looks the way it does.
Most teaching presents an algorithm as a finished artefact, states what it does, shows the update rule and moves on. He works forwards from the problem instead. Here is what we want, here is the obvious approach, here is why it fails, here is the modification that fixes it, and now the thing on the slide looks inevitable rather than arbitrary. Once you have seen kernels derived that way, or understood the bias variance decomposition as a consequence rather than a formula to memorise, that understanding does not leave.
The scope is the classical canon done properly. Nearest neighbours, perceptron, linear and logistic regression treated seriously, the bias variance tradeoff, kernels and support vector machines, Gaussian processes, decision trees, bagging and boosting. This is the material that the deep learning wave pushed out of fashion and that turns out to matter enormously the moment you are working with tabular data, limited samples, or a problem where you need to explain your model to somebody. Gradient boosting still wins a great many practical problems, and understanding why requires exactly this material.
The lecture notes on the course site are strong enough to use on their own, and the recordings are real classroom sessions with students asking questions, which is more valuable than it sounds because the questions are almost always the one you were about to have. There is a texture to a live lecture with genuine confusion in the room that a scripted studio recording cannot reproduce. What you have to accept is the date. These are 2018 recordings, which in this field is a long time ago, and the practical consequence is that modern deep learning is essentially absent.
There is no transformer content, nothing on large language models, and the deep learning coverage that exists is brief and now dated. That is not a flaw in the course, it is a fact about when it was recorded, but it does mean this cannot be your only source. Pair it with something current, whether that is the fast.ai course, Karpathy's series or a modern university deep learning course, and treat this as the foundation layer underneath. The prerequisites are also real.
This was taught to Cornell undergraduates who had done the mathematics, and the pace reflects that. You need linear algebra and probability at working fluency, and if you do not have them the lectures will move past you quickly and pleasantly without you noticing you have stopped following. The public version also has no assessment, since the assignments and exams stayed with the enrolled students, so you have to construct your own checks by implementing what you have watched. That is the right thing to do anyway, and it is worth saying explicitly, because watching thirty hours of excellent lectures produces a very convincing sensation of competence that evaporates the first time you sit in front of an empty file.
The production quality is what you would expect from a camera at the back of a lecture theatre, and while the audio is fine, this is not a polished product and does not try to be. My 4.5 is one of the higher ratings in this catalogue and I am comfortable with it. It costs nothing, it requires no account, and it teaches the foundations better than most paid courses teach anything. The deductions are for age and for the absence of any structured practice, both of which you can work around, and neither of which touches the central point, which is that this is what excellent teaching sounds like.