Training Machine Learning Models with Ising Machines

Sayantan Pramanik, Kaumudibikash Goswami, Sourav Chatterjee, M Girish Chandra

Keywords: Ising machines, trust region optimisation, training ML models, second order optimsation

TL;DR: In this study, we have used Ising machines to help train machine learning models. This is in contrast to typical applications of the Ising machines in combinatorial optimisation and as ML models.

Abstract: 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.

Submission Number: 33