Chapter 1 gave you the model — a file of numbers. Chapter 2 gave you its raw material — a table of examples. This chapter connects them: the patient loop that turns a table into a model, the choice of columns that decides what can be learned at all, the honesty ritual that keeps everyone from fooling themselves, the central disease every practitioner fears, and the two comparisons that give any score its meaning.
5 topics
"We trained a model" is the sentence the whole field stands on, and this is the chapter where it stops being a black box. The mechanism turns out to be almost disappointingly simple: guess, measure how wrong the guess was, nudge the model's numbers to be slightly less wrong, and repeat — thousands upon thousands of times. No studying, no flash of insight. A loop.
Five topics, one arc. First the loop itself, walked once and slowly, because it is machine learning. Then the columns — what the model gets to look at, and how to tell a genuine clue from noise and from the one kind of column that cheats. Then the train/test split, the ritual that makes honest measurement possible at all. Then the book's heart: overfitting — a model memorizing its homework instead of learning the subject — and generalization, the thing we actually wanted. And finally the question every score must answer before it means anything: good enough compared to what?
Training in one circle — the loop this chapter walks