22w5055: Interpretability in Artificial Intelligence

The Banff International Research Station will host the "Interpretability in Artificial Intelligence" workshop in Banff from May 1 - 6, 2022. State-of-the-art deep learning networks are achieving strong predictive power, but the gain in accuracy often comes at the price of transparency, meaning that the prediction of the model is not interpretable. These so-called black-box models raise critical challenges in high-stake domains, such as healthcare, crime recidivism, or finance, where a wrong decision can have very harmful consequences. For instance, a tool to risk-stratify patients trained on a very unbalanced datasets can assign all new cases to the most prevalent risk-category, failing hence to identify clinical factors that might separate high- and low-risk patients. In other cases, ethnic, economic, or social factors, which might by chance correlate with a patient group, might be wrongly used by the model to classify new patients.