Simon Fraser University

Published purpose

Even in the era of big data, many domains may still suffer from lack of high quality labeled data for machine learning. Thus, zero-shot transfer learning is particularly important, which explores methodologies to transfer models learned from a supervision-rich domain to a domain with no properly labeled training data. In this project, we will focus on zero-shot transfer learning for sophisticated models and from a supervision-rich domain to a domain whose relationship is even unknown. We will tackle several data science challenges and develop principled methods. Particularly, we will investigate cross-domain data augmentation methods, which help to generate training or testing data for an unknown domain, or transfer data from a supervision-rich domain to an unknown domain. The techniques developed in this project may be used in many important problems, such as multi-lingual natural language processing (NLP) tasks and multi-modal learning.