Developing Trustworthy AI: Dr. Agus Sudjianto on High-Risk Machine Learning

On April 30th, 2024, Dr. Agus Sudjianto, the former Executive Vice President and Head of Model Risk for Wells Fargo, presented his seminar “Performance is Not All You Need: Developing and Validating High Risk Machine Learning,”  at UNC Charlotte as a part of The Center for TAIMing AI’s seminar series. In this seminar, Dr. Sudjianto discussed how performance and accuracy alone are insufficient for machine learning models, especially in high-risk industries, and should be supplemented by other metrics like risk assessment and interpretability.

In recent years, the adoption of machine learning models has skyrocketed; however, this rapid adoption has been accompanied by a string of high-profile model failures, ultimately leading to increased scrutiny over the reliability of AI. Dr. Sudjianto’s presentation stressed the importance of model validation, one of the cornerstones of model risk management. This safeguard has two key components: conceptual soundness and outcome analysis. Conceptual soundness involves examining data quality, model assumptions, and interpretability. Outcome analysis assesses performance under real-world conditions where the operational environment deviates from the training data.

Finally, Dr. Sudjianto’s seminar called for a proactive model risk management approach. This approach allows for the transformative potential of AI while maintaining trustworthiness and safety in AI models.