← AI Terminology
Curriculum Learning
Curriculum learning is a training strategy where examples are presented to a model in a meaningful order — typically from easy to hard — mimicking the way humans learn, rather than training on random shuffles of the data.
Introduced by Bengio et al. (2009), it often improves convergence speed and final model quality.
Introduced by Bengio et al. (2009), it often improves convergence speed and final model quality.
Why It Matters in AI
Random data shuffling ignores the difficulty structure of the training set. Starting with clear, easy examples lets the model learn basic patterns before being exposed to ambiguous or hard ones — analogous to teaching multiplication before calculus. Curriculum strategies are used in LLM pre-training (sequence length curriculum), RL (reward shaping), and data quality filtering (train on high-quality data first).
Key Points
| Aspect | Description |
|---|---|
| Self-paced | Model decides what's hard based on its own current loss — dynamic curriculum |
| Data quality | Train on clean, high-quality data first; introduce noisier web data later |
| Easy-to-hard | Sort examples by estimated difficulty; present simpler ones first |
| RL curriculum | Start with simple task variants; gradually increase complexity as agent improves |
| Difficulty proxy | Sentence length, loss on a reference model, human annotation confidence, curriculum heuristics |
| Seq-length sched | LLM training: start with short sequences, gradually increase — speeds up early training |
Simple Analogy
A piano teacher doesn't start students with Rachmaninoff — they begin with scales, then simple pieces, then progressively harder repertoire. Each stage builds the foundation for the next. Curriculum learning applies the same pedagogy to neural network training.
Common Usage Examples
- LLM pre-training: GPT and Llama training often starts with 2K token sequences, scaling to 32K+ later
torchdata.dataloader2with custom samplers that order examples by precomputed difficulty scores- Self-paced learning:
loss_mask = (per_sample_loss > threshold)— skip examples the model already handles well - RL: OpenAI Five (Dota 2) used curriculum of game modes from simple to full 5v5 matches
- Speech recognition: train on clean read speech before adding conversational and noisy data
Summary
In short: Curriculum learning presents training examples in order from easy to hard — accelerating convergence by building a foundation before tackling the hard cases.