Continual Learning

Continual learning (also called lifelong learning or incremental learning, depending on the community) is the machine-learning paradigm in which a model is exposed to a stream of tasks or data distributions over its lifetime, rather than being trained once on a fixed dataset. The defining challenge is catastrophic forgetting: standard gradient-based training on a new task tends to overwrite the parameters that supported prior tasks, causing accuracy on earlier tasks to collapse. Continual-learning research designs training procedures, architectures, and replay mechanisms that mitigate this trade-off so that performance on the full task sequence improves monotonically rather than degrading over time.

Approaches are typically grouped into three families: (i) regularization-based methods that penalize changes to parameters deemed important for prior tasks (for example, Elastic Weight Consolidation, Synaptic Intelligence); (ii) replay- or rehearsal-based methods that maintain a buffer of stored or generated examples from past tasks and interleave them with new training; and (iii) architectural or parametric-isolation methods that allocate dedicated sub-networks or capacity to each task. The term "continual learning" consolidated in its modern machine-learning sense through Parisi et al.'s 2019 review "Continual Lifelong Learning with Neural Networks," which unified usage across the prior lifelong-learning, incremental-learning, and never-ending-learning literatures. The paradigm has gained renewed urgency with large language models: training-time continual learning (continued pre-training on new corpora) and inference-time continual learning (in-context adaptation across a session) are now active research directions and the basis of several recent industry bets, including Sutskever-era Safe Superintelligence's reported first model.