Zero shot adaptation and learning
Massimiliano Mancini (Ph.D. Student)
Domain Adaptation is a transfer learning scenario where the goal is to build a model addressing a task, e.g. classification, in a target domain with no or few images are labeled. Given a large amount of labeled data in a domain, i.e. the source, with a different input distribution from the target, e.g. synthetic to real images, solving this task requires tackling the so-called domain shift problem. Despite recent advances, there are two fundamental limitations of standard domain adaptation techniques: source and target label spaces need to be (at least partially) shared and unlabeled target domain data is used to model the shift between the domains. As a consequence, those techniques can only recognize seen categories in seen target domains. Recognizing unseen classes without any labeled training sample is the goal of Zero-Shot Learning. However, standard zero-shot learning techniques do not consider domain shift among training and test domains, e.g. photo vs sketch. At the same time, some recent Domain Generalization and Predictive Domain Adaptation methods tackle domain shift problem without requiring target domain data, and by exploiting the presence of multiple source/auxiliary domains to disentangle domain specific features and semantic features. However, these approaches assume that the label spaces in different domains are shared. In this project, we propose to merge these two worlds and build models capable of recognizing unseen classes as in Zero-Shot Learning and in unseen domains as in Predictive Domain Adaptation and Domain Generalization.
|Primary Host:||Barbara Caputo (Politecnico di Torino & Italian Institute of Technology)|
|Exchange Host:||Zeynep Akata (University of Tübingen)|
|PhD Duration:||01 November 2016 - 31 October 2020|
|Exchange Duration:||01 March 2020 - 30 June 2020 - Ongoing|