Making learning more transparent using conformalized performance prediction

As machine learning systems become increasingly entangled with human decision-making processes, it is of social, ethical, and economic importance to be able to understand and explain the limits of what can be said about such technologies. We need to be able to formulate meaningful questions about the behavior or performance of learning systems that are accessible to a diverse audience, and we must have the tools required to provide unambiguous answers to these questions. From the viewpoint of responsible and sustainable system design, the tools used to answer questions about the learning system are arguably even more important than the algorithms underlying the learning system itself. In this context, there is an important line of research related to conformal prediction, a general-purpose methodology for attaching high-probability guarantees to data-driven predictors [25, 16, 9, 8]. For example, if the user specifies a confidence level of say 90% in advance, questions such as "I've trained a classifier. Given a new instance, what labels seem most likely?"