Ising machines are promising for neural network training and combinatorial optimization. However, their slow statistical sampling hinders neural network training efficiency. This paper introduces a universal concept for ultrafast statistical sampling using analog Ising machines by injecting noise. Experiments with an opto-electronic Ising machine demonstrate accurate Boltzmann distribution sampling and unsupervised neural network training, matching software-based methods in accuracy. Simulations show orders-of-magnitude faster sampling than software methods, enabling Ising machines for machine learning and other applications beyond combinatorial optimization.

Fabian Böhm, Diego Alonso-Urquijo, Guy Verschaffelt, Guy Van der Sande