Integrating Deep Learning and Bayesian Reasoning

Deep learning (DL) is an excellent function estimator which has amazing result on perception tasks such as visualization recognition and text recognition. But, its inner architecture acts as a black box, because the users cannot understand why such decisions are made. Bayesian reasoning (BR) provides explanation facility and causal reasoning in terms of uncertainty which is able to overcome demerit of DL. This paper is to propose a framework for the integration of DL and BR by leveraging their complementary merits based on their inherent internal architecture. The migration from deep neural network (DNN) to Bayesian network (BN) involves extracting rules from DNN and constructing an efficient BN based on the rules generated, to provide intelligent decision support with accurate recommendations and logical explanations to the users.