Quantum Computing-Driven Optimization of Machine Learning Algorithms for Embedded Systems

Abstract

The development of quantum technologies opens new possibilities for the optimization of machine learning (ML) algorithms. In this contribution, we report on the experimental use of quantum computing for training ML algorithms and models designed to be deployed for inference on embedded systems with limited computational resources. Due to the small size of resource-efficient models for embedded systems, this application of quantum computing provides an excellent opportunity to evaluate a technology that is currently being developed in the noisy intermediate-scale quantum era. We develop an approach based on transforming classical supervised learning problems into a discrete form, enabling their solution using quantum annealing (QA) methods. Three models are presented: a binary XNORNet network, a support vector machine (SVM) with a discrete representation of dual coefficients, and a convolutional neural network. All experiments are performed on the MNIST dataset, using classical optimization methods, simulated annealing, and QA on the D-Wave Advantage system. While classical gradientbased methods maintain superior absolute accuracy (up to 95.97% for SVM), the D-Wave system demonstrates a significant reduction in core optimization time, achieving speedups of up to 15.6× compared to classical optimization on a central processing unit when the system overhead is excluded. Preliminary results suggest that while QA entails a slight accuracy and memory requirements penalty, a drastic reduction in active annealing duration serves as a promising strategy to reduce training latency in future embedded-focused ML applications.