continual learning

Continual learning enables models to learn new tasks sequentially without forgetting previous ones. **Challenge**: Standard training on new data causes catastrophic forgetting. Model faces stability-plasticity trade-off. **Approaches**: **Regularization-based**: EWC (Elastic Weight Consolidation) penalizes changes to important weights, SI (Synaptic Intelligence) tracks parameter importance during training. **Replay-based**: Store examples from previous tasks (experience replay), generate synthetic samples of old tasks. **Architecture-based**: Progressive networks add new modules, PackNet prunes and freezes subnetworks per task, modular networks with task-specific routing. **For LLMs**: Continual pre-training on new domains, instruction tuning without losing base capabilities, mixing old and new data. **Evaluation**: Forward/backward transfer metrics, average accuracy across all seen tasks. **Applications**: Models that learn over time in production, personalization without forgetting, adapting to distribution shift. **Current research**: Rehearsal-free continual learning, continual RLHF, efficient memory management. Critical for deploying AI systems that improve over time without expensive retraining.