Domain Adaptation
A technique in machine learning where a model trained in one domain is adapted to work in a different but related domain.
Definition
Addresses distribution shifts between source (training) and target (deployment) domains through methods like feature-alignment, adversarial domain classifiers, or fine-tuning on small target-domain labeled samples. Proper governance includes benchmarking adapted models on held-out target data and ensuring no degradation in critical subgroups.
Real-World Example
A speech-recognition model trained on U.S. English accents is adapted for U.K. English using only 10 hours of British-accent recordings. Engineers apply adversarial domain adaptation to align feature spaces, improving word-error rate by 30% on U.K. test sets without retraining from scratch.
.svg)
