The guide
Before the networks, the models you can still reason about. Linear regression has a closed form, so you can solve it exactly and then watch gradient descent arrive at the same answer, which makes the optimiser much less mysterious.
The through line is the bias-variance tradeoff. Ridge regression adds a penalty that makes the fit worse on the training set on purpose, and better on everything else. k-fold cross validation exists because a single train/test split gives you a number with error bars you cannot see. Both are about the same anxiety: the score you measured is not the score you will get.
Scaling comes first for a reason. Distance-based methods and gradient descent both quietly assume your features are comparable, and neither warns you when they are not.