ABSTRACT:

Machine learning systems, in virtually every deployed system, encounter data from a qualitatively different distribution than what they were trained upon.  Effectively dealing with this problem, known as domain shift, is thus perhaps the key challenge in deploying machine learning methods in practice.  In this talk, I will motivate some of these challenges in domain shift, and highlight some of our recent work on two topics.  First, I will present our work on determining if we can even just evaluate the performance of machine learning models under distribution shift, without access to labelled data.  And second, I will present work on how we can better adapt our classifiers to new data distributions, again assuming access only to unlabelled data in the new domain.