Picture a student who, while learning Spanish, suddenly erased all the French they had mastered. This is exactly what happens to neural networks: catastrophic forgetting (or catastrophic interference) is a model's tendency to abruptly lose skills acquired on earlier tasks when trained on new data.
Why it happens
A network encodes all its knowledge in a single set of shared weights. When trained on a new task, gradient descent adjusts those weights to minimize only the new loss, with no memory of the old one. The parameters that encoded the previous task get overwritten.
Formally, at step $t$ we minimize only:
$$\theta^* = \arg\min_\theta \; \mathcal{L}_{\text{new}}(\theta)$$
whereas we would also like to preserve $\mathcal{L}_{\text{old}}(\theta)$. With no constraint, nothing protects the old knowledge.
How to mitigate it
Several families of strategies exist in continual learning:
| Approach | Principle |
|---|---|
| Rehearsal | Replay old examples during new training |
| Regularization (e.g. EWC) | Penalize changes to weights deemed important |
| Dynamic architectures | Allocate new parameters per task |
- Regularization adds a term that gently "freezes" critical weights.
- Fine-tuning a large language model is a frequent victim: overly aggressive tuning degrades the model's original general capabilities.
Learning without forgetting remains one of AI's great open challenges: the balance between plasticity (learning the new) and stability (retaining the old).