Reasoning over Bayesian Networks using Semantic Artificial Neural Networks

Representation of application domains, related concepts and their dependencies is often achieved using Bayesian Networks. In Bayesian Networks nodes represent random variables and arcs represent their dependencies. Since inference over Bayesian Networks is a complex task in this work a novel approach for representing and reasoning over Bayesian Networks using Semantically labeled Neural Networks is proposed and evaluated. Using Semantic Neural Networks combines advantages of Neural Networks such as wide adoption and highly optimized implementations while preserving the interpretability of Bayesian Networks which is an important requirement, especially in medical applications. In addition the proposed approach is evaluated over medical datasets with positive results.