continual

Definition

A setting where training never stops: new data or new tasks arrive in sequence and must be absorbed by an already-trained model. The defining obstacle is catastrophic forgetting — naive further training on new material rapidly erases what came before. What distinguishes it from simply retraining on everything is the assumption that past data is unavailable or too expensive to revisit, and that assumption is what makes it hard.

Explainers using this term

• Paper Walkthrough: Macaron-V1 — A Frozen Base plus a Mixture of LoRAs, Built to Keep Learning After LaunchMacaron-V1: Towards Open Continual Learning with Self-Improvement and Mixture-of-LoRA