M achine learning is a powerful tool for transforming data into computational models that can drive user-facing applications. However, potential users of such applications, who are often domain experts for the application, have limited involvement in the process of developing them. The intricacies of applying machine-learning techniques to everyday problems have largely restricted their use to skilled practitioners. In the traditional applied machine-learning workflow, these practitioners collect data, select features to represent the data, preprocess and transform the data, choose a representation and learning algorithm to construct the model, tune parameters of the algorithm, and finally assess the quality of the resulting model. This assessment often leads to further iterations on many of the previous steps. Typically, any end-user involvement in this process is mediated by the practitioners and is limited to providing data, answering domain-related questions, or giving feedback about the learned model. This results in a design process with lengthy and asynchronous iterations and limits the end users' ability to affect the resulting models.Consider the following case study of machine-learning practitioners working with biochemists to develop a protein taxonomy by clustering low-level protein structures (Caruana et al. 2006). The project lead recounted their experience in an invited talk at the Intelligent User Interfaces's 2013 Workshop on Interactive Machine Learning (Amershi et al. 2013). First, the practitioners would create a clustering of the protein structures. Then, they would meet with the biochemists to discuss the results. The biochemists would critique the results (for example, "these two proteins should / should not be in the same cluster" or "this cluster is too small"), providing new constraints for the next iteration. Following each meeting, the practitioners would carefully adjust the clustering parameters to adhere to the given con- (Cohn, Caruana, and McCallum 2003;Caruana et al. 2006). These algorithms were intended to give people the ability to rapidly iterate and inspect many alternative clusterings within a single sitting.Their later approach is an example of interactive machine learning, where learning cycles involve more rapid, focused, and incremental model updates than in the traditional machine-learning process. These properties enable everyday users to interactively explore the model space and drive the system toward an intended behavior, reducing the need for supervision by practitioners. Consequently, interactive machine learning can facilitate the democratization of applied machine learning, empowering end users to create machine-learning-based systems for their own needs and purposes. However, enabling effective end-user interaction with interactive machine learning introduces new challenges that require a better understanding of end-user capabilities, behaviors, and needs.This article promotes the empirical study of the users of interactive machine-learning systems as a metho...

How can end users efficiently influence the predictions that machine learning systems make on their behalf? This paper presents Explanatory Debugging, an approach in which the system explains to users how it made each of its predictions, and the user then explains any necessary corrections back to the learning system. We present the principles underlying this approach and a prototype instantiating it. An empirical evaluation shows that Explanatory Debugging increased participants' understanding of the learning system by 52% and allowed participants to correct its mistakes up to twice as efficiently as participants using a traditional learning system.

Abstract-Research is emerging on how end users can correct mistakes their intelligent agents make, but before users can correctly "debug" an intelligent agent, they need some degree of understanding of how it works. In this paper we consider ways intelligent agents should explain themselves to end users, especially focusing on how the soundness and completeness of the explanations impacts the fidelity of end users' mental models. Our findings suggest that completeness is more important than soundness: increasing completeness via certain information types helped participants' mental models and, surprisingly, their perception of the cost/benefit tradeoff of attending to the explanations. We also found that oversimplification, as per many commercial agents, can be a problem: when soundness was very low, participants experienced more mental demand and lost trust in the explanations, thereby reducing the likelihood that users will pay attention to such explanations at all.

Many machine-learning algorithms learn rules of behavior from individual end users, such as taskoriented desktop organizers and handwriting recognizers. These rules form a "program" that tells the computer what to do when future inputs arrive. Little research has explored how an end user can debug these programs when they make mistakes. We present our progress toward enabling end users to debug these learned programs via a Natural Programming methodology. We began with a formative study exploring how users reason about and correct a text-classification program. From the results, we derived and prototyped a concept based on "explanatory debugging", then empirically evaluated it. Our results contribute methods for exposing a learned program's logic to end users and for eliciting user corrections to improve the program's predictions.

No abstract