Catastrophic Forgetting

Catastrophic forgetting, also called catastrophic interference, is the tendency of a neural network to lose previously learned knowledge when it is trained on a new task. Because the same set of weights encodes everything the model knows, the updates that reduce error on the new task can overwrite the representations that supported the old one. A network that learned to recognize cats and dogs may forget how to tell them apart after later being trained only on birds.

The stability-plasticity dilemma

At the heart of the problem is a trade-off researchers call the stability-plasticity dilemma. A model needs plasticity to absorb new information, but too much plasticity destroys old knowledge; it needs stability to retain what it learned, but too much stability prevents new learning. Continual learning, sometimes called lifelong or incremental learning, is the field devoted to keeping both in balance.

Why it matters for self-hosted models

For anyone fine-tuning a model repeatedly on their own data, catastrophic forgetting is a practical risk: each round of training on a new corpus can erode the general capabilities the base model arrived with. Mitigation strategies include replaying samples of old data alongside new, freezing parts of the network, using parameter-efficient adapters that leave the base weights untouched, and penalizing changes to weights deemed important for prior tasks. These approaches let a sovereign operator extend a model without quietly degrading it.

For related concepts, see our entries on overfitting, the model checkpoint for preserving good states, and synthetic data.