Curriculum Learning

What is curriculum learning?

Nobody learns math by starting with olympiad problems. You count, then add, then multiply. Curriculum learning brings that "easy first" intuition into machine learning: instead of feeding the model data in random order, you sort it from easy to hard and let it climb step by step.

How is it different from random shuffling?

Random feeding: easy to get lost
If hard examples flood in from the start, the model can get confused, take forever to converge, or even give up and just learn the easiest patterns.
Curriculum feeding: build the foundation first
Let the model absorb simple examples and learn the basic rules before adding difficulty. Convergence is steadier and the final result is often better.

How do you decide difficulty?

By hand
Label examples by experience: sentence length, image clarity, number of steps in a task.
Let the model choose
Some methods measure each example's loss or uncertainty and focus next on what the model still struggles with — like a teacher adjusting the pace on the fly.

Where is it used?

Training LLMs
Feed short simple text first, then longer and harder reasoning, to give the model a smooth start.
Robot learning
From simple grasps to complex chains of tasks, ramping up gradually.
Computer vision
Start with clear, easy examples, then introduce occlusion and blur.

What are its limits?

Curriculum learning isn't a silver bullet. If your difficulty ordering is bad — or the task has no real easy-to-hard scale — the gains can vanish. It's a scheduling strategy that shines alongside good data and a good model, not a replacement for them.

Bottom line: curriculum learning arranges training data easy-to-hard, so the model learns steadily and lands better results.