The thermodynamic approach to neural networks begins with the work of Lars Onsager, who solved the transition matrices for simple Ising systems. A milestone was reached in the 70's with the work of Ilya Prigogine, who used nonlinear nonequilibrium thermodynamics to understand reaction-diffusion systems. Along the way, there was a considerable amount of work in probability theory, like the Ornstein-Uhlenbeck equation that describes mean-reverting generators.