Title: Model-Agnostic explanations for Machine Learning predictions
Advisor: Carlos Guestrin
Supervisory Committee: Carlos Guestrin (Chair), Jevin West (GSR, I School), Jeffrey Heer, and Sameer Singh (UC Irvine)
Abstract: Machine learning is at the core of many recent advances in science and technology. Unfortunately, the important role of humans is an oft-overlooked aspect in the field. As machine learning becomes a crucial component of an ever-growing number of user-facing applications, there are multiple reasons to focus on interpretable machine learning. A possible approach to interpretability in machine learning is to be model-agnostic, i.e. to extract post-hoc explanations by treating the original model as a black box.
In this work, we present the thesis that model-agnostic interpretability, despite its challenges, is useful in the human-machine learning interaction in a variety of domains / tasks. The goals of this work are threefold: (1) to develop methods that produce model-agnostic explanations, (2) to investigate and evaluate how humans react to such explanations, and (3) to develop methods to incorporate human feedback into models, thereby “closing the loop”. Model-agnosticism inherently leads to some approximation, as the only way to have perfect faithfulness in explanation is by inspecting the black-box model itself. A major contribution of our work is considering and measuring how humans interact with model-agnostic explanations, and designing them in a way that optimizes such interaction.