Residual Sum-Product NetworksFabrizio Ventola; Karl Stelzner; Alejandro Molina; Kristian Kersting
In: Manfred Jaeger; Thomas Dyhre Nielsen (Hrsg.). Proceedings of the 10th International Conference on Probabilistic Graphical Models. International Conference on Probabilistic Graphical Models (PGM-2020), September 23-25, Aalborg, Denmark, Pages 545-556, Proceedings of Machine Learning Research, Vol. 138, PMLR, 2020.
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.