ModelRefs / Continual Learning — AI Glossary
Continual Learning — AI Glossary
Training a model on a sequence of tasks or data distributions without catastrophic forgetting of previously acquired knowledge.
Overview
Continual learning (CL) methods—elastic weight consolidation, progressive neural networks, experience replay, LoRA-based parameter isolation—aim to incrementally update models as new data arrives. Relevant to maintaining model freshness for RAG-free knowledge updates and domain adaptation in production.
Reference details
| Topic | training |
|---|---|
| Also known as | lifelong learning, incremental learning |
| Last reviewed | 2026-06-24 |
Related terms
How it works
- A model already trained on one distribution is updated on new data, which ordinarily overwrites the weights that encoded the old behaviour.
- Regularisation approaches estimate which weights matter for prior tasks and penalise moving them, so capacity is protected rather than preserved by accident.
- Replay approaches interleave a sample of old data with the new, which is the most reliable mitigation and requires keeping that data available.
- Isolation approaches freeze the base and add new parameters per task, so nothing can be forgotten because nothing is overwritten.
- Evaluation must cover the old tasks as well as the new, since a model that improved on the new distribution and collapsed on the old has not learned continually.
Example: Forgetting is measured on what you stopped testing
A model scores 80% on task A. Fine-tuned on task B it reaches 85% on B, and nobody re-runs A. Measured, A has fallen to 55% — a 25-point loss invisible in the release note because the suite only covered the new task. Interleaving even 10% of task A data during the B fine-tune typically recovers much of that gap, at a 10% increase in training data. The cost of the mitigation is small; the cost of not measuring is that the regression is silent.
Commonly confused with
Often conflated with fine-tuning, which is the mechanism rather than the problem. Fine-tuning updates a model on new data; continual learning is the discipline of doing so repeatedly without destroying what came before, and its defining concern is catastrophic forgetting. It is also not retrieval: RAG adds knowledge at inference without touching weights, which sidesteps forgetting entirely and is usually the cheaper answer when the goal is freshness rather than new capability.
When to use it
Reach for it when:
- A model must absorb new capability repeatedly over time rather than once.
- Prior-task performance is part of the acceptance criteria, not an afterthought.
- Retrieval cannot supply what is needed because the gap is capability rather than knowledge.
Reach for something else when:
- The need is current facts, where retrieval is cheaper and carries no forgetting risk.
- Old tasks are not evaluated, in which case forgetting will happen unobserved.
- A single adaptation is required, which is ordinary fine-tuning.
Continue your research
Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Continual Learning — AI Glossary.
Frequently asked questions
What is Continual Learning?
Training a model on a sequence of tasks or data distributions without catastrophic forgetting of previously acquired knowledge.
Is Continual Learning the same as lifelong learning?
Yes — lifelong learning, incremental learning are common aliases for Continual Learning.
What concepts are related to Continual Learning?
Closely related concepts include catastrophic forgetting, domain adaptation, transfer learning.