Transfer Learning

This problem deals with how to transfer knowledge accumulated through machine learning to new tasks, to solve or learn to solve them faster. Knowledge transfer is a natural phenomenon in human beings, for instance, when basic algebra learned in junior classes helps us quickly learn calculus in the later years. We gradually build our knowledgebase and learn to solve increasingly complex tasks, on the basis of our earlier experiences. Unfortunately, most of machine learning has traditionally focused on problem oriented approaches (sort of "horses for courses") disregarding the usefulness of one approach in a related but different problem that the learner may face later. Transfer learning bridges this gap.

Open questions include how to structure cumulative knowledge so that it can be readily applied in later tasks, how to select portions of the knowledgebase that will be most effective for a new task, how to map elements of an old problem to those of a new problem, and how to measure the amount of transfer that a given approach accomplishes so that we can compare different approaches. This project attempts to develop the theories to answer these questions. Check out a workshop on this topic, that I co-chaired.