Unsupervised domain adaptation in the presence of a disjoint label space
We often have data from two domains that share the same space but have different distributions in it. An example of this is photographs of hand-written digits and photographs of digits from street signs of the same height, width and channels. One of these domains (called the source domain) has labels that we can use to train a supervised model, while the other domain (called the target domain) is unlabelled. The problem is to utilise the labelled information from the source domain to aid the unsupervised training of the target domain. The specific problem we have chosen to address is the case when the target domain has labels not previously encountered in the source domain. For this, we use ideas from zero-shot learning and unsupervised domain adaptation.
Keywords: zero shot learning, GANs, deep learning