Replay Buffer
Also called Experience Replay
A stash of old examples the model keeps revisiting so it doesn’t forget them while learning new ones.
Think of it like
Mixing a few old flashcards into every study session so you don’t lose what you learned last month.
Example
While teaching a model new topics, the trainer sprinkles in old data from the buffer so past skills stay sharp.
How it actually works
A replay buffer stores past experiences or examples and re-injects samples of them into ongoing training. In continual learning it preserves old knowledge; in reinforcement learning it also breaks correlations between consecutive samples, making updates more stable. The design questions are what to keep and how to sample from it.
For product teams
A practical guardrail so updating the model doesn’t silently degrade the things it already did well.
For engineers
A memory of prior samples replayed during training to stabilize learning and counter forgetting.
Related
- Catastrophic forgetting — A tool against catastrophic forgetting.
- Continual Learning — Common in continual learning.
- Reinforcement learning — Also core to reinforcement learning.
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