Representing Uncertainty with a New Type of Stochastic Neural Networks

Many interesting complex systems are stochastic. In order to model such complex systems, much ongoing research is looking at how to precisely model uncertainty in performance. In this paper, we proposed a novel type of stochastic neural network (SNN), in which dynamic features are added to the input layer allowing any non-deterministic system to be modeled. The SNNs capture randomness from the additional input nodes fed with internal random signals. These random signals, combined with weights between the additional nodes and the hidden nodes, allow stochastic output even though the network is deterministic. To validate this approach, a preliminary experiment was performed. To show the SNN’s basic ability to represent uncertainty, a SNN model is trained to represent a model of beta distribution. Experiments verify the basic feasibility of the approach.