Reading Machine Learning Without Getting Lost
Dense books become easier when you separate the argument from the algebra on the first pass.
When a machine learning chapter becomes difficult, the temptation is to slow down immediately and parse every symbol. Sometimes that is necessary. Often, it is too early.
Map before memorizing
On the first pass, I try to answer three questions: what problem is being solved, what assumption makes the solution possible, and what changes after the solution is introduced. This creates a map before the details arrive.
Use code as a second language
A short implementation can translate notation into behavior. Even incomplete code can clarify which values are fixed, which values are learned, and where the data flows.
Return with questions
The second pass is where the math becomes worth reading closely. You are no longer staring at symbols in isolation. You are asking the text to explain something you have already tried to build.