We are interested in understanding what a machine learning system could ask a user to improve its performance and when it is appropriate to ask the user. Our goal is to improve learning while minimizing user costs. We are exploring representations of a system’s internal state and determining when a system could use input to improve performance