Abstract
We address the problem of causal effect estimation where hidden confoundersare present, with a focus on two settings: instrumental variable regressionwith additional observed confounders, and proxy causal learning. Our approachuses a singular value decomposition of a conditional expectation operator,followed by a saddle-point optimization problem, which, in the context of IVregression, can be thought of as a neural net generalization of the seminalapproach due to Darolles et al. [2011]. Saddle-point formulations have gatheredconsiderable attention recently, as they can avoid double sampling bias and areamenable to modern function approximation methods. We provide experimentalvalidation in various settings, and show that our approach outperforms existingmethods on common benchmarks.