Divide et Impera quantum neural networks for modular hybrid computing architectures

Abstract

Quantum machine learning is rapidly emerging as a promising and potentially more sustainable approach that can complement traditional, energy-hungry HPC, AI, and GPU resources-particularly for the demanding global challenges the world urgently needs to address. In recent years, the first prototypes of quantum processors have reached the market, are sometimes available via the cloud, and distributed quantum computing is beginning to move from a largely theoretical concept to a demonstrated reality. However, these devices are still noisy and qubit-limited. Therefore, it is desirable to design quantum machine learning models with a reduced quantum register and a limited number of quantum gates, avoiding overly deep and wide quantum circuits. Here, we explore this idea by proposing a novel divide-et-impera-based training protocol that splits quantum neural networks into smaller components and better governs their optimization. We benchmark our approach on real data and real-world prediction tasks in intertwined energy and environmental domains, such as the status of charging stations for electric vehicles and the global air quality index. Our results suggest a design principle for quantum machine learning algorithms across a wide range of applications, which can be more feasibly tested on state-of-the-art quantum processors or within modular quantum architectures being developed for future hybrid infrastructures interconnected via high-speed communication links.