End-User Debugging of Machine-Learned Programs: Toward Principles for Baring the Logic

Abstract

Many applications include machine learning algorithms intended to learn “programs” (rules of behavior) from an end user’s actions. When these learned programs are wrong, their users receive little explanation as to why, and even less freedom of expression to help the machine learn from its mistakes. In this paper, we develop and explore a set of candidate principles for providing salient debugging information to end users who would like to correct these programs. We informed the candidate principles through a formative study, built a prototype that instantiates them, and conducted a user study of the prototype to collect empirical evidence to inform future variants. Our results suggest the value of exposing the machine’s reasoning process, supporting a flexible debugging vocabulary, and illustrating the effects of user changes to the learned program’s logic.