Conditional sum-product networks: Modular probabilistic circuits via gate functions

While probabilistic graphical models are a central tool for reasoning under uncertainty in AI, they general not as expressive deep neural models, and inference is notoriously hard slow. In contrast, such sum-product networks (SPNs) capture joint distributions ensure tractable inference, but still lack the power of intractable based on networks. this paper, we introduce conditional SPNs (CSPNs)—conditional density estimators multivariate potentially hybrid domains—and develop structure-learning approach that derives both structure parameters CSPNs from data. To harness (DNNs), also show how to realize by conditioning vanilla input using DNNs gate functions. contrast whose high-level can be explicitly manipulated, naturally used building blocks modular maintains interpretability. experiments, demonstrate competitive with other yield superior performance structured prediction, estimation, auto-regressive image modeling, multilabel classification. particular, employing encoders decoders within variational autoencoders help relax commonly mean field assumption turn improve performance.