When distributions differ
A model trained on one source domain often degrades on a target domain with a different distribution, such as a model trained on news text applied to medical notes. Domain adaptation adjusts the model so it performs well on the target despite this distribution shift.
Labeled versus unlabeled targets
- Supervised adaptation fine tunes on a small set of labeled target examples.
- Unsupervised adaptation has only unlabeled target data and aligns the model to the target distribution without target labels.
The amount of available target labels shapes which strategy fits.
The idea
Common techniques
- Continued pretraining on unlabeled target text adapts the representations before any task fine tuning.
- Feature alignment encourages source and target representations to look similar, so a source classifier transfers.
- A small amount of labeled target data for fine tuning is often the simplest effective fix when available.
Key idea
Domain adaptation closes the gap from distribution shift by aligning representations or fine tuning on the target, with the method chosen by how much labeled target data exists.