Abstract
In Bayesian filtering, states and parameters of probabilistic state-space models are inferred in an online manner. Using the Free Energy Principle, the state-space model is cast to a generative model p and the posterior distributions of interest are approximated using recognition distributions or beliefs q. The factorisation of state-space models into state transitions and observation likelihoods over time supports forming a factor graph and performing inference via message passing. Tools for message passing on factor graphs typically employ a scheduling procedure, in which a separate algorithm or compiler takes the model description and returns which nodes should pass messages where at what time. This can be sufficiently expensive to form a bottleneck. Moreover, it’s not a biologically plausible mechanism for governing message passing. I explore the possibility of passing messages without a scheduler. A designated terminal node should pass an initial message, which will arrive at an initial variable. The corresponding belief is updated, a local Free Energy is computed and the belief is emitted to neighbouring factor nodes. From there on out, whenever an updated belief arrives at a factor node, the node fires messages to all other variables if the local Free Energy surpasses a threshold. If not, the node becomes silent.
