Adapter
A small trainable module slipped between a frozen model’s layers to teach it a new skill.
Think of it like
Like plug-in expansion cards for a computer — the machine stays the same; you add a card for the new capability.
Example
You insert adapter modules into a frozen base model and train just those to specialize it for medical text.
How it actually works
Adapters add small bottleneck layers inside the network and train only them while the base stays frozen, making adaptation cheap and modular. Different adapters can be swapped in for different tasks on the same shared model. LoRA is a close relative that adapts existing weights rather than inserting new layers; both fall under PEFT.
For product teams
Modular skills for one base model — mix and match without retraining the whole thing.
For engineers
Inserted bottleneck modules trained while the backbone is frozen; a PEFT approach, kin to LoRA.
Related
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