Updates optimized for the new distribution can overwrite parameters important to older capabilities. Mixing old data, regularizing changes, replay, or modular adaptation can help preserve prior behavior.
Catastrophic forgetting is the loss of previously learned behavior when a model is trained on new data or tasks.
Updates optimized for the new distribution can overwrite parameters important to older capabilities. Mixing old data, regularizing changes, replay, or modular adaptation can help preserve prior behavior.