Training Machine Learning Models with Ising Machines

In this study, we use Ising machines to help train machine learning models by employing a suitably tailored version of opto-electronic oscillator-based coherent Ising machines with clipped transfer functions to perform trust region-based optimisation with box constraints. To achieve this, we modify such Ising machines by including non-symmetric coupling and linear terms, modulating the noise, and introducing compatibility with convex-projections. The convergence of this method, dubbed $i$Trust has also been established analytically. We validate our theoretical result by using $i$Trust to optimise the parameters in a quantum machine learning model in a binary classification task. The proposed approach achieves similar performance to other second-order trust-region based methods while having a lower computational complexity. Our work serves as a novel application of Ising machines and allows for a unconstrained optimisation problems to be performed on energy-efficient computers with non von Neumann architectures.