Presentation Abstract:

Faced with unlabeled data in deployment that is sampled from a different distribution than that which generated the training data, all bets are off. Moreover, while numerous heuristics have been proposed for this vague setting, it remains unclear when any among them are applicable. One way to render these problems identifiable is to impose some (assumed) causal structure, both over how the variables are related to each other, which factors are potentially manipulable (and, complementarily, which are domain-invariant). Unlike conventional problems in causality, where the goal is to estimate the effect of a manipulation (a change in the policy for prescribing the treatment), here the manipulation has already happened, and our goal is to leverage the causal structure to adapt our predictors appropriately. In this talk, I will discuss some structures under which these problems are identifiable and some of the challenges (and solutions) for applying these ideas in deep learning settings.