Catastrophic Forgetting in Machine Learning (2026 Guide)

Catastrophic forgetting is the sharp drop in performance on previously learned tasks after a model is trained on new tasks. In neural networks, shared parameters are updated for the new task and can overwrite internal representations that were important for earlier tasks.

It happens because a single set of weights is used for many tasks. Gradient updates for the newest data minimise its loss, even if that hurts older tasks. Overfitting to recent data, interference between tasks, and online learning on drifting data distributions all contribute to catastrophic forgetting.

Common strategies include weight regularisation, Elastic Weight Consolidation, rehearsal or replay buffers, careful transfer learning, and ensemble or modular architectures. In practice, you often combine several of these methods and monitor validation performance per task to detect and control forgetting.