Residual Sum-Product Networks

Tractable yet expressive density estimators are a key building block of probabilistic machine learning. While sum-product networks (SPNs) offer attractive inference capabilities, obtaining structures large enough to fit complex, high-dimensional data has proven challenging. In this paper, we present a residual learning approach to ease the learning of SPNs, which are deeper and wider than those used previously. The main trick is to ensemble SPNs by explicitly reformulating sum nodes as residual functions. This adds references to substructures across the SPNs at different depths, which in turn helps to improve training. Our experiments demonstrate that the resulting residual SPNs (ResSPNs) are easy to optimize, gain performance from considerably increased depth and width, and are competitive to state of-the-art SPN structure learning approaches. To combat overfitting, we introduce an iterative pruning technique that compacts models and yields better generalization.