Boltzmann machine neural network devices using single-electron tunnelling

Abstract

We proposed a method of implementing the Boltzmann machine neural network on electronic circuits by making use of the single-electron tunnelling phenomenon. The single-electron circuit shows stochastic behaviour in its operation because of the probabilistic nature of the electron tunnelling phenomenon. It can therefore be successfully used for implementing the stochastic neuron operation of the Boltzmann machine. The authors developed a single-electron neuron circuit that can produce the function required for the Boltzmann machine neuron. A method for constructing Boltzmann machine networks by combining the neuron circuits was also developed. The simulated-annealing operation can be performed easily by regulating an external control voltage for the network circuits. A sample network was designed that solves an instance of a combinatorial optimization problem. Computer simulation demonstrated that, through the simulated-annealing process, the sample network can converge to the global minimum energy state that represents the correct solution to the problem.