Catastrophic forgetting
When training a model on something new makes it abruptly lose skills it already had.
Think of it like
Like cramming for a new language so hard that last year’s subject leaks out of your head entirely.
Example
You fine-tune a model heavily on legal text and suddenly it’s worse at basic math it used to handle easily.
How it actually works
Neural networks store knowledge in shared weights, so updating them for a new task can overwrite the patterns that supported old ones. It’s a central obstacle to sequential learning and aggressive fine-tuning. Mitigations include mixing old data back in, freezing parts of the network, using PEFT so the base stays intact, or regularizing weights to resist large changes.
For product teams
Why fine-tuning can quietly degrade a model elsewhere — always test the skills you didn’t train on.
For engineers
Overwriting of prior representations during new-task training; countered by rehearsal, freezing, PEFT, or regularization.
Related
- Continued pretraining — A danger of continued pretraining.
- Fine-tuning — Also a risk in fine-tuning.
- PEFT — Mitigated by PEFT.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome