Maryland team benchmarks tunable quantum neural network on ion and superconducting hardware

A University of Maryland, College Park team led by Djamil Lakhdar-Hamina reported a neural network whose inference step runs on two different quantum computers, testing whether quantum neural networks can deliver practical gains, phys.org reported. The study appears in Physical Review Letters (DOI 10.1103/9bp2-42v3), with an arXiv version also cited (DOI 10.48550/arxiv.2507.21222). The researchers trained the network classically, then ran decision-making on quantum hardware that can be tuned from a purely classical mode to a fully quantum mode with more measurement uncertainty. They ran the same network on trapped-ion and superconducting platforms and tested handwritten-digit recognition. Moderate quantum uncertainty improved accuracy versus the fully classical mode, before additional noise hurt reliability. For some images the classical mode misclassified, the quantum version sometimes produced the correct label, which the team linked to hardware noise. Extra canceling operations helped compare how the two architectures responded to noise. phys.org said the results suggest noise in quantum machine learning is not only a flaw to remove and can, when structured carefully, push decisions in a useful direction. The work is framed as hardware benchmarking as quantum systems grow larger.