A theory of learning from different domains
Develops a theory of domain adaptation, bounding a classifier's target error via source error and a measurable divergence between domains.
This work develops theory for domain adaptation, where a classifier trained on a source domain must perform well on a differently distributed target domain with little or no labeled data. It bounds target error in terms of source error and a classifier-induced divergence measure that can be estimated from finite unlabeled samples. Assuming a hypothesis performs well in both domains, these quantities characterize target error. It also bounds the error of models minimizing a convex combination of source and target errors, showing how to choose the optimal weighting.
Based on: A theory of learning from different domains · Machine-mediated learning
Curated by Aramai Editorial
Read summary →