Encyclopedia · 176 concepts

Core ML Concepts · advanced · concept 31 of 176

Continual Learning & Catastrophic Forgetting

Neural networks trained on a new task tend to overwrite what they knew, which is why models are retrained from scratch instead of updated in place. Continual learning studies how to add knowledge without erasing the old, one of the most practical unsolved problems in the field.

Key terms

Catastrophic forgettingReplay bufferElastic weight consolidationTask interferenceOnline learning

Where you meet it in the real world

Personalization without retraining, robots that learn on the job, keeping deployed models current