What Is Catastrophic Forgetting in AI Models?
Teaching a model something new can sometimes make it suddenly, badly worse at something it already knew — a well-known failure mode in continued training.
Learning something new, losing something old
Catastrophic forgetting is what happens when further training a model on a new task or dataset causes it to sharply lose performance on tasks it previously handled well — the new training overwrites the parameter patterns the old capability depended on, rather than adding to them cleanly.
Why it's a particular risk during fine-tuning
It shows up most often when fine-tuning a general-purpose model on a narrow, specialized dataset — the model can become excellent at the new narrow task while noticeably degrading on the broad general capability it started with, if the fine-tuning data is too narrow or training runs too long.
How it's typically mitigated
Common defenses include mixing some of the original general-purpose training data back in during fine-tuning, using a lower learning rate so updates are gentler, and stopping training as soon as the target task improves without over-training past that point.
Frequently Asked Questions
What is catastrophic forgetting?
A failure mode where further training a model on a new task causes it to sharply lose performance on a previously learned task, because the new training overwrites the parameter patterns the old capability relied on.
How can catastrophic forgetting be prevented?
Common approaches include mixing original training data back in during fine-tuning, using a gentler learning rate, and stopping training as soon as the new task is learned rather than continuing past that point.