ICML 2011 workshop on unsupervised and transfer learning

Towards Heterogeneous Transfer Learning
Qiang Yang
Hong Kong University of Science and Technology, Clearwater Bay, Kowloon, Hong Kong
and
Guirong Xue and Yong Yu
Shanghai Jiao Tong University, Shanghai, China

Transfer learning aims to learn new concepts for a learning task by reusing knowledge from related but different domains. Most existing transfer learning tasks have focused on knowledge transfer between domains with the same or similar feature representation spaces. However, the potential of transfer learning should stem from its ability to acquire knowledge from very different feature spaces. We call these transfer learning tasks heterogeneous transfer learning. In this article, we highlight some examples of a heterogeneous transfer learning via knowledge transfer between text and images and between domains without any explicit feature mappings. In image learning problems, such as image classification and clustering, relatively few labeled data are available to train a high quality model. Our idea is to exploit the similarity of the domains at a deeper level to link them together, in order to transfer knowledge from textual domains. For example, we ask: is it possible to supplement a image classifier with a large quantity of unlabeled text to improve its performance? When two domains do not have any explicit feature mappings, is it still possible to find a mapping between them? We review three of our recent works in answering these questions, and show that by carefully crafting a translator via resources such as Flickr, it is possible to transform knowledge from text domains to image problems. We also illustrate how transfer learning can be done between two very different domains even when no translator can be found between them, by identifying and maximizing the commonalities among the structures of the different domains.