Personalizing Machine Learning Systems with Explanatory Debugging
KEC 1003 Monday, January 5, 2015 - 4:00pm to 4:50pm Todd Kulesza School of EECS Oregon State University Abstract: How can end users efficiently influence the predictions that machine learning systems make on their behalf? Traditional systems rely on users to provide examples of how they want the learning system to behave, but this is not always practical for the user, nor efficient for the learning system. This talk explores a different personalization approach: a two-way cycle of explanations, in which the learning system explains the reasons for its predictions to the end user, who can then explain any necessary corrections back to the system. In formative work, we study the feasibility of explaining a machine learning system’s reasoning to end users and whether this might help users explain corrections back to the learning system. We then conduct a detailed study of how learning systems should explain their reasoning to end users. We use the results of this formative work to inform Explanatory Debugging, our explanation-centric approach for personalizing machine learning systems, and! present an example of how this novel approach can be instantiated in a text classification system. Finally, we evaluate the effectiveness of Explanatory Debugging versus a traditional learning system, finding that explanations of the learning system’s reasoning improved study participants’ understanding by over 50% (compared with participants who used the traditional system) and participants’ corrections to this reasoning were up to twice as efficient as providing examples to the learning system. Speaker Bio: Todd Kulesza recently completed his PhD in Computer Science at Oregon State University. His research interests are in human interactions with intelligent systems, with a focus on helping end users control such systems efficiently and effectively. He was a co-organizer of the ACM IUI 2013 Workshop on Interactive Machine Learning, and his work in this area has received an ACM CHI Best Paper award in 2014 and Honorable Mention in 2012.
_______________________________________________ Colloquium mailing list [email protected] https://secure.engr.oregonstate.edu/mailman/listinfo/colloquium
