← AI Terminology
Continual Learning
Continual learning (lifelong learning) trains models on a stream of tasks or data distributions over time while retaining prior competence.
It contrasts with one-shot i.i.d. training then freeze.
It contrasts with one-shot i.i.d. training then freeze.
Why It Matters in AI
The real world is non-stationary. Products must absorb new tools, policies, and domains without full retrain-from-scratch. Continual learning research and practical adapter/replay strategies address this.
Key Points
| Aspect | Description |
|---|---|
| Evals | Forward/backward transfer metrics |
| Goals | Plasticity for new + stability for old |
| Methods | Replay, regularisation, isolation (masks/adapters), expansion |
| Related | Catastrophic forgetting, domain adaptation |
| Hard problem | No perfect free lunch yet at foundation-model scale |
| LLM practice | Adapter stacks, periodic mixture retrain, RAG for facts |
Simple Analogy
A professional who keeps learning new tools across a career without forgetting how to do last year’s job.
Common Usage Examples
- Sequential task benchmarks (Split-CIFAR, etc.)
- LoRA per skill, compose at inference
- Replay buffer of prior domains
- RAG instead of finetuning for fast-changing facts
Summary
In short: Continual learning updates models over time without erasing past skills — lifelong adaptation under non-stationary data.