H.B. Keller Colloquium Series

Despite their many successes, modern machine learning systems can behave in surprising ways when deployed in real-world contexts. A key challenge is that the standard ML pipeline is built on criteria that are domain- and application-agnostic, and these procedures provide few guarantees beyond aggregate performance on the data distribution that the model was trained on. In this talk, I will review some of these difficulties, and discuss approaches that leverage ideas and formalism from causal inference to control or adjust the signal that models use to make predictions. I will focus specifically on how causal formalism can be used to design robust and adaptable ML systems in the presence of "spurious" associations.