1 paper · 1 filter
Yuli Slavutsky, Yuval Benjamini
Zero-shot learning methods typically assume that the new, unseen classes encountered during deployment come from the same distribution as the the classes in the training set. Howev…