catastrophic forgetting

Catastrophic forgetting, also called catastrophic interference, occurs when an artificial neural network trained sequentially on new data or tasks overwrites the weights that encoded earlier ones, causing a sharp drop in performance on what it learned before. Unlike gradual forgetting in humans, the loss can be severe even after brief additional training.

It is a central problem in continual learning, where a model must learn a series of tasks over time, and it also shows up when large language models fine-tuned for a specific task lose some of their general abilities. Common mitigations include replaying part of the earlier data, regularization methods that limit changes to important weights, and training only small added modules while the base model's weights stay frozen.

This entry is based on AIPOST articles and widely known facts. If something is wrong, please send us a correction request.