The goal of domain adaptation [4] is to address a given task (e.g. classification) in a given domain (i.e. the target) exploiting labeled data in a different domain (i.e. the source). Standard DA approaches have two fundamental limitations: 1) target data must be available and 2) the task and its label space must be shared among the source and target domains. For the former, recent domain generalization (DG) and predictive domain adaptation (PDA) methods, showed that we can solve DA even without target data. For the second limitation, despite progresses on open and partial DA there is currently no method that can perform DA when source and target do not share the same task unless target samples and associated labels are available Focusing on classification, recognizing unseen classes without any labeled training sample is the goal of Zero-Shot Learning (ZSL). In this project, I will attempt at merging these two worlds, building models capable of recognizing unseen classes, as in zero-shot learning, in unseen domains, as in predictive Domain Adaptation and Domain Generalization.