← AI Terminology
Catastrophic Forgetting
Catastrophic forgetting is when a neural network trained on a new task rapidly loses performance on previously learned tasks.
It is the central challenge of continual learning.
It is the central challenge of continual learning.
Why It Matters in AI
Fine-tuning an LLM on a narrow domain can erase general skills. Mitigations (replay, regularisation, LoRA isolation, mixture methods) matter for lifelong products and sequential domain adaptation.
Key Points
| Aspect | Description |
|---|---|
| Eval | Measure old and new tasks after each stage |
| Cause | Shared weights overwritten by new gradients |
| Related | Continual learning, transfer learning |
| Symptom | Old-task accuracy collapses after new training |
| LLM angle | Domain finetunes vs base capabilities |
| Mitigations | Replay buffers, EWC, progressive nets, adapters/LoRA |
Simple Analogy
Cramming for a new exam so hard you forget last semester’s material — limited mental real estate overwritten.
Common Usage Examples
- Before/after MMLU when domain-tuning
- Replay mix of old data during finetune
- Task-specific LoRA adapters instead of full overwrite
- EWC-style regularisers in continual setups
Summary
In short: Catastrophic forgetting is wiping old skills when learning new ones — the core risk of sequential fine-tuning.