10-732 Robustness and Adaptation in Shifting Environments
12 units, Carnegie Mellon University
Under distribution shift, the source domain which produces labeled training data and the target distribution which produces unlabeled deploymenttime data might diverge. In general, if one can assume nothing about how the past relates to the future, then all bets are off. However, under a variety of reasonable settings, we might hope for robustness andor adaptation. Descriptions of such settings in the literature vary in key ways twodomain discrete shift setups versus continual shifts benign shifts, versus targeted adversarial attacks, shifts that preserve factors of causal structure andor support versus shifts that respect geometric invariances. In this class we will take a grand tour of the literature on robustness and adaptation under distribution shift, bringing together perspectives from distributional robustness, structural causal models, adversarial learning in both senses, and the wild west of heuristics and benchmarks that have been proposed in the deep learning literature. Students will be responsible for submitting short responses to each weeks readings, and presenting one paper and leading the associated discussion.
Prerequisites: none listed
Courseweave needs JavaScript to show sections, workloads and planning tools.