Steps Toward Robust Artificial Intelligence

AI technologies are being integrated into high-stakes applications such as self-driving cars, robotic surgeons, hedge funds, control of the power grid, and weapons systems. These applications need to be robust to many threats including cyberattack, user error, incorrect models, and unmodeled phenomena. This talk will survey some of the methods that the AI research community is developing to address two general kinds of threats: The "known unknowns" and the "unknown unknowns". For the known unknowns, methods from probabilistic inference and robust optimization can provide robustness guarantees. For the unknown unknowns, the talk will discuss three approaches: detecting model failures (e.g., via anomaly detection and predictive checks), employing causal models, and constructing algorithm portfolios and ensembles.  For one particular instance of model failure---the problem of open category classification where test queries may involve objects belonging to novel categories---the talk will include recent work with Alan Fern and my students on providing probabilistic guarantees.