Inference by replication - a schematic description
The main achievement of the project are in devising a general distributed inference method for tackling problems that can be mapped onto densely connected graphs, where evidential data nodes are discrete. The method has been applied to the hard computational problems of Ising perceptron training, lossy compression, inference in instances of the Sherrington-Kirkpatrick model and the inverse Ising model; all of which are relevant to both application domains and theoretical research. We showed how such methods can be used for the various problems, gaining better understanding of their properties, their advantages and drawbacks, and the related computational complexity. In a separate study we also investigated the dynamics of non-equilibrium systems held at different temperatures, but coupled through cross-system interactions. We studied the stationary solutions obtained and their stability.