Continual Learning refers to methods where models continuously learn new tasks or data without losing previously acquired knowledge – a problem known as "Catastrophic Forgetting."
Typical strategies include regularization methods that protect important parameters for previous tasks (e.g., Elastic Weight Consolidation), replay methods that reuse old data or their representations, as well as architectural approaches that create separate subnetworks for new tasks.
Continual Learning is especially relevant for applications where data arrives sequentially and a model should be permanently adapted without retraining from scratch.
The challenges lie in mastering the balance between stability (retaining old knowledge) and plasticity (adapting to new information).
Research areas also include lifelong learning and multi-task learning, which are closely related to Continual Learning.