Approximated Probabilistic Inference on a Dynamic Bayesian Network Using a Multistate Neural Network
Makito Oku · IEICE Proceeding Series · 2014
Dynamic Bayesian networks (DBNs) are flexible tools for modeling complex relationship among time-evolving random variables. An application of DBNs to computational neuroscience is to represent the internal model, which the brain uses to simulate the environment, as a DBN. The exact inference on the DBN defines the optimal behavior for both sensory and motor signal processing. However, since the exact inference requires huge computational resources, approximation methods are the key to utilize the DBN representation. Here, I propose a new heuristic algorithm for probabilistic inference on the DBN using a multistate neural network. Each random variable of the DBN is replaced by a multistate neuron, and the directional links of the DBN are translated into the nonlinear interactions among the multistate neurons. To approximate backward dependencies among variables in the DBN, the network supports a bottom-up error-reporting mechanism against top-down predictions. The proposed method is tested on a simple partially observable Markov decision process task, and exhibits better performance than ancestral sampling method.