Deep Networks with Factor Graph in Reduced Normal Form paradigm

A Probabilistic Graphical Model describes, through a graph, the independent relations among its stochastic variables. The Factor Graphs are a type of Graphical Model and, in their Reduced Normal Form (FGrn), they allow to design the Probabilistic Systems like block diagrams. The completely local learning provides to FGrn paradigm with a great modularity and flexibility. The inference is done letting the evidence, injected in the network at an arbitrary point, propagate in bi-directional way using the Belief Propagation algorithm.

The paradigm can manage, naturally, variables of different types becoming an important tool for the information fusion.

PhD Thesis - Slides