Meta-learning is currently used to solve few-shot learning problems that include learning new categories within one speciﬁc sub-domain or dataset (Finn et al., 2017; Snell et al., 2017; Vinyals et al., 2016; Sung et al., 2018). It would be signiﬁcantly more practical if the model could easily learn new categories in new unseen sub-domains, especially ones considerably different from the training sub-domain(s). This area is largely unexplored, although a variety of cross-domain few-shot learning approaches have been recently proposed (Kang & Feng, 2018; Triantaﬁllou et al., 2018; Choi et al., 2019).
Preliminary work has also been done as part of the author’s MScR thesis on DomainNet dataset (Peng et al., 2019a), which includes sub-domains such as real-world images, cliparts or sketches. Our model would be meta-trained on a variety of sub-domains, and then it would be supposed to learn new categories from a new unseen sub-domain, using only very few examples. The goal is to develop methods that are more suitable for this cross-domain setting than the existing meta-learning methods.