Date(s) : 16/12/2013 iCal
10 h 55 min - 11 h 55 min
Domain Adaptation of Weighted Majority Votes via Perturbed Variation-based Self-Labeling\n\nBy Emilie Morvant\, IST Austria.\n\nTalk at Friiam/Frumam\, 2e étage\, St Charles\n\nIn machine learning\, the domain adaptation problem arrives when the test (target) and the train (source) data are generated from different distributions. A key applied issue is thus the design of algorithms able to gen- eralize on a new distribution\, for which we have no label information. We consider the specific PAC-Bayesian situation focused on learning classification models defined as a weighted majority vote over a set of real- valued functions. In this context\, we present PV-MinCq a new framework that general- izes a non-adaptative algorithm (MinCq). PV-MinCq follows the next principle. Jus- tified by a theoretical bound on the tar- get risk of the vote\, we provide to MinCq a target sample labeled thanks to a per- tubed variation-based self-labeling focalized on the regions where the source and target marginals appear similar. We also study the influence of our self-labeling\, from which we deduce an original process for tuning the hy- perparameters. Our experiments show very promising results on a synthetic problem.
Catégories Pas de Catégories