Learning Bounds for Domain Adaptation [M27]
John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, Jenn Wortman University of Pennsylvania Setup: Plentiful but biased source data and relevant but scarce target data
Example task: identifying genes in DNA sequences Domains: different organisms source target

We develop a uniform convergence bound that is computable from finite labeled & unlabeled samples Minimize convex combinations of source & target risk Generalizes standard single-domain bounds as well as previous domain adaptation bounds