Distributionally Robust Learning of Sum-Product Networks

Rohith Peddi, Vibhav Giridhar Gogate

Keywords: Distributionally Robust Optimization, Sum-Product Networks

TL;DR: An efficient approach for learning distributionally robust SPNs

Abstract: Sum-Product networks (SPNs) are generative probabilistic models that use a deep architecture comprised of alternating layers of sum and product nodes to compactly represent a high-dimensional joint probability distribution. In this paper, we consider the problem of learning robust SPNs from the lens of distributionally robust optimization (DRO) under the Wasserstein metric. We show that SPNs learned by maximizing likelihood exhibit poor performance when data is subject to noise/corruptions. To address this issue, we construct probabilistic uncertainty sets and leverage the tractability of SPNs to efficiently learn distributionally robust SPNs. We show our proposed approach's efficacy on a collection of benchmark datasets.