Learning that never stops
Continual learning trains a model on a sequence of tasks that arrive over time, while keeping performance on earlier tasks. The central tension is the stability plasticity dilemma: the model must stay stable enough to remember yet plastic enough to learn new things.
Families of methods
- Replay stores or generates examples from past tasks and mixes them into new training so old knowledge is rehearsed.
- Regularization penalizes changes to weights that were important for earlier tasks, slowing drift where it would hurt.
- Architectural methods add new parameters for new tasks while protecting old ones.
The cycle
Why it is hard
Without care, each new task causes catastrophic forgetting. But naive remedies can also limit how much the model adapts. Good continual learning balances the two, often combining a little replay with regularization so the model accumulates skills rather than trading one for another.
Key idea
Continual learning trains on a stream of tasks while balancing stability and plasticity, using replay, regularization, or growth to retain old skills as it gains new ones.