Kikuchi free energies with weak consistency constraints: change point learning in switching linear dynamical systems

Exact inference in probabilistic models is often infeasible due to (i) a complicated conditional independence structure, and/or (ii) troublesome local integrals. Most challenging inference problems found in physics, such as the computation of the partition function in an Ising model or Boltzmann machine are examples of problems that suffer from a complex structure. All variables are binary, but the cycles in the model prevent an efficient recursive formulation of an inference algorithm.