Title:  Building data efficient models using active learning and similarity comparisons

Committee: 

Dr. Davenport, Advisor 

Dr. Dyer, Chair

Dr. Pananjady

Abstract: The objective of the proposed research is to develop methods for learning data efficient models based on learning from similarity comparisons and active learning. The motivation behind using comparisons stems from the idea that it is easier to compare different objects than evaluate every object on an absolute basis. Active learning reduces the label complexity of tasks by selecting the optimal objects to be labelled. We apply these methods to exhibit improved performance in problems such as metric learning - which involves learning a representation of objects via comparisons of distances between them, classification - where we formulate label acquisition as a problem of soliciting response to similarity queries and preference learning - where the goal is to estimate a user based on their item preferences. Further, we consider the problem of learning representations from unlabelled data where the ultimate goal is to perform well on some downstream tasks. Here, we are interested in understanding the conditions under which good performance in the former step ensures good results in the downstream tasks.