curriculum learning

Schools don't open with calculus on day one — they start with counting, then arithmetic, then algebra, building up. Curriculum learning brings that ordering to machine training: instead of throwing examples at the model in random order, you present easy examples first and gradually work up to the hard ones, the way a good teacher sequences a lesson.

Ordinary training shuffles the data so the model sees an unpredictable mix every step. Curriculum learning deliberately shapes the order — simple, clean, unambiguous examples early, then increasing difficulty, noise, or rarity. The intuition is that early easy examples help the model find a sensible region of solutions before it tackles confusing cases that might otherwise send it down a bad path. You need a notion of "difficulty" to sort by — sentence length, image clarity, how many other examples it resembles, or the model's own error on each one.

It can speed up training and sometimes reach a better final result, especially on hard problems where a random order leaves the model floundering early. But the evidence is mixed: for many ordinary tasks a well-tuned random-order setup does just as well, and a clumsy curriculum can hurt by starving the model of variety or by your difficulty ranking being wrong. Treat it as a sometimes-helpful trick, not a law — and remember the hardest part is often just deciding what "easy" and "hard" mean.