Wasserstein Source-Class Subset Selection for Transfer Learning under Label Shift
A method for few-shot transfer learning when the source and target label sets differ arbitrarily (extreme, possibly disjoint, label shift). Rather than assume shared classes, it uses the Wasserstein distance between class-conditional feature distributions to select and reweight a subset of source classes whose combined distribution is closest to the target, and justifies this selection as minimizing an upper bound on target generalization error. It also introduces a distribution-distance-based augmentation for the few-shot target data.
Class-based Subset Selection for Transfer Learning under Extreme Label Shift
Existing work within transfer learning often follows a two-step process -- pre-training over a large-scale source domain and then finetuning over limited samples from the target domain. Yet, despite …