Stability-Plasticity Dilemma
Overview
Continual learning studies how models can adapt to changing tasks or environments without losing useful knowledge from the past. Common approaches include replaying earlier examples, regularizing important parameters with methods such as elastic weight consolidation, isolating task-specific parameters, and expanding model capacity; reinforcement learning also uses consolidation across timescales and adaptive architectures. These methods are evaluated on settings such as class- and domain-incremental vision tasks, medical image segmentation, and non-stationary control, with measures that include retention, new-task learning, efficiency, and generalization. No approach resolves every trade-off: replay can burden resources or representations, while protecting old knowledge can limit adaptation, so evaluation beyond forgetting remains important.
Latest papers
19 papers3 papers in the last four weeks, against 1 the four weeks before. 0.0% of all new papers.