These are notes from a university course: regression and its variants, classification, generative models, support vector machines, learning theory, unsupervised methods and a section on reinforcement learning. All with the mathematics in front of you. You read them with a pencil or you do not read them.
What I took from it
The bias and variance diagnosis, which is the most useful thing any material on this subject has given me. When a model is no good there are two distinct causes with opposite remedies: either it is too simple for the problem, or it memorised the data it saw. Telling them apart decides whether you need more data — expensive and slow — or a different approach — cheap. Without that diagnosis, the conversation always ends in asking for more data, which is the comfortable answer.
The second is understanding what a library is doing when you call a three-word function. You do not need to derive it every time, but having derived it once changes how you read results and how quickly you get suspicious of a number that looks too good.
And the third: regularisation as a way of telling a model to prefer the simple. It is the same principle I apply when designing a solution: between two that explain the same thing, the simpler one maintains itself.
What I left out
Almost everything operational, because it is not there: no dirty data, no costs, no deployment, no client waiting. And the emphasis belongs to another era. What survives intact is the underlying statistics, which is what almost nobody revisits today and what separates people who use tools from people who understand what the tools hand back.