As machine learning is used more frequently and for higher risk applications, there is a growing desire to be able to explain how a model makes its predictions rather than simply treating it as a predictive black-box. The field of explainable machine learning has emerged to address this desire. However, there have been two consistent and related critiques of the field: (1) Specificity - "explainability'" is not a monolithic concept and, as a result, we must define which specific aspect of a model’s behavior an explanation captures and how capturing that behavior is useful for a specific application and (2) Rigor - there is a lack of rigorous protocols for evaluating explainability methods.
As a result, we aim to demonstrate how to design rigorous explainability pipelines for specific applications. To do so, we use three motivating applications for explainability: helping end-users interact with a model, helping domain-experts use a model to discover new knowledge, or helping model-developers debug a model. Along the way, we demonstrate that Global Counterfactual Explanations are often a useful tool. In the proposed work, we plan to increase the usefulness of Global Counterfactual Explanations for Model Debugging.
Ameet Talwalkar (Chair)
Carlos Guestrin (University of Washington)
Been Kim (Google Brain)
Marco Tulio Ribeiro (Microsoft Research)
Zoom Participation. See announcement.