|Institution:||Hong Kong University of Science and Technology|
|Keywords:||Machine learning; Heterogeneous computing|
|Full text PDF:||https://doi.org/10.14711/thesis-991012554867303412;http://repository.ust.hk/ir/bitstream/1783.1-88029/1/th_redirect.html|
Artificial Intelligence has been enjoying an unprecedented boom recently. The huge success, however, still heavily depends on massive labeled data. Transfer learning, leveraging knowledge from a source domain to improve predictive models in a target domain which does not have sufficient labeled data, has been more and more popular. In transfer learning, a source and a target domain have a discrepancy in any of the feature space, distribution, label space, and predictive models, which traditional machine learning algorithms cannot handle. The majority of existing transfer learning algorithms focus on homogeneous transfer learning where the feature space, the label space as well as the predictive model are shared. However, homogeneous transfer learning algorithms lose their power if the shared feature space is insufficient to build satisfactory predictive models, or if a source domain in the same feature and label space cannot be found. In this case, heterogeneous transfer learning (HTL) is desired. Provided with a source domain in a completely different feature space or label space, heterogeneous transfer learning algorithms transfer knowledge in different perspectives, just as we human beings with a multi-sensory system are capable of transferring lip reading knowledge to improve speech understanding if one is whispering in a very low voice. The key to transfer learning is to building either instance-based or feature-based mappings between a source and a target domain. In this thesis, we focus on developing scalable and principled methodologies to build feature mappings under different heterogeneity: 1) how to build semantic correspondence between a pair of source and target domains in different feature spaces to pave the way for successful knowledge transfer; 2) how to transfer when a source and a target domain have different feature spaces, while each domain has either single type of data or multiple types of data; 3) how to transfer when a source and a target domain have different label spaces. We have applied our algorithms to multiple large-scale real-world datasets from different applications including computer vision, social media, health care, and urban computing. This thesis introduces this research frontier and points out some promising research issues for extensive investigation.