Boltzmann Machine Using Superconducting Circuits

We study the design and optimization of the Boltzmann machine hardware using superconducting circuits as a new stochastic information processing method. The Boltzmann machine is an artificial neural network of stochastic binary models wherein the energy function is determined by the given set of parameters, and the output is obtained by stochastic state transitions of the system to the energy stable states according to the Boltzmann distribution. By adjusting the set of parameters, arbitrary functions can be embedded in energy-stable states, which have applications in data dimensionality reduction and generative models. The hardware of a Boltzmann machine using superconducting circuits consists of quantum flux parametrons (QFPs), one of the superconducting circuits, magnetically coupled to each other. In this study, we designed the Boltzmann machine hardware in which logic gates such as NOR are embedded in energy-stable states. Furthermore, we applied maximum likelihood estimation (MLE), machine learning method, as an operating-point optimization method, and confirmed the effectiveness of this method in experiments.