The issue of area generalization issues studying predictive fashions which can be strong to distribution shifts when deployed in new, beforehand unseen environments. Present strategies usually require labeled knowledge from a number of coaching environments, limiting their applicability when labeled knowledge are scarce. On this work, we research area generalization in an anti-causal setting, the place the end result causes the noticed covariates. Underneath this construction, setting perturbations that have an effect on the covariates don’t propagate to the end result, which motivates regularizing the mannequin’s sensitivity to those perturbations. Crucially, estimating these perturbation instructions doesn’t require labels, enabling us to leverage unlabeled knowledge from a number of environments. We suggest two strategies that penalize the mannequin’s sensitivity to variations within the imply and covariance of the covariates throughout environments, respectively, and show that these strategies have worst-case optimality ensures beneath sure lessons of environments. Lastly, we exhibit the empirical efficiency of our method on a managed bodily system and a physiological sign dataset.
† Apple
‡ ETH Zürich
