תאור הקורס:
Supervised machine learning relies on the fundamental assumption that data is
sampled iid from the same distribution at train time and at test time. But in virtually
any realistic application, this assumption is unlikely to hold. In this seminar we will
survey papers that study when, how, and why learning algorithms (such as ERM) can
fail when the assumption is violated. We will study various failure modes that stem
from different reasons underlying why train and test distribution can differ, including:
natural distribution drift, model-induced distribution shift, adversarial manipulation of
inputs, and strategic behavior of self-interested users.
Topics: (list of papers will be published towards the beginning of the semester)
- Out-of-distribution (OoD) generalization
- Distribution shift and drift
- Semi-supervised domain adaptation
- Covariate shift and debiasing
- Inverse propensity weighing (IPW)
- Invariant representation and risk minimization (IRM)
- Causality vs. (spurious) correlation
- Decision-dependent distribution shift
- Adversarial learning and robustness
- Strategic classification
- Performative prediction
- Distributional robustness