Causally Motivated Robustness to Shortcut Learning in Machine Learning

Watch a 45-minute research lecture from the Simons Institute where Maggie Makar from the University of Michigan explores novel techniques for addressing machine learning model robustness under distribution shifts. Learn about shortcut learning as a major failure mode in robust machine learning, where models develop unstable associations that break down when test distributions change. Discover causally-motivated regularization approaches designed to prevent shortcuts by enforcing models to follow specified causal structures. Examine scenarios both where potential shortcuts are known in advance and where such prior knowledge is not available. Gain insights into practical challenges faced when deploying machine learning models that need to maintain performance across different distributions.