A novel approach to ML learning by taking the dual of the log-partition function, resulting in a saddle-point (or adversarial) problem with kinetic variables. Reviews were consistent that the approach was novel and may spur further research.