Agus Sudjianto (Executive Vice Present and Head of Corporate Model Risk at Wells Fargo) discusses ‘Explainable AI’, a topic we explore with the backdrop of the US agencies’ recent Request for Information. The US agencies (the Federal Reserve, FDIC, the OCC, the CFBP, and the National Credit Union Administration) sought industry comment on how to support the responsible adoption of Artificial Intelligence and Machine Learning technologies, with broad topic coverage including explainability, model governance, data sources, the potential for model bias and the role of third parties.
Agus explains the difference between post-hoc explainability techniques and building explainability into a design feature or constraint. He articulates how these approaches each fit with different types of applications, while also giving the important reminder that models are designed to estimate and approximate, and must be understood in that context. We also discuss the 10-year anniversary of SR11-7, the US supervisory guidance on model risk, and its important legacy worldwide.