Everyone can recite that a flexible model fits the training data better and generalises worse. Far fewer can predict when it is about to happen to them, because the two errors are usually described rather than watched. The block below puts both on screen and lets you cause the problem deliberately, which is the fastest way to recognise it later in something you care about.
Place points, then raise the complexity. Training error falls to zero while the held-out error climbs.
Click in the field to place points of the colour you are placing, then switch colours and place some of the other. Four or so of each is enough for the two error numbers to start meaning something.
You describe the goal in your own words and it asks questions back until the plan is one you would actually follow. These are the shapes that goal usually takes.
Not a fixed syllabus — the plan is built around your goal and cuts what does not serve it. These are the topics it draws from, and whatever you get wrong comes back until it stops coming back.
The mechanics are the same whatever the subject — a slider is a slider whether it is moving a coefficient or a rate of return. What changes is what it is a slider for.
Or see how it works and what it costs.