First Quantum Neural Network Implemented Across Two Computing Platforms

Researchers have implemented a quantum neural network across two different quantum computing platforms, trapped ions and superconducting qubits, a step toward testing whether such systems can live up to their theoretical promise. The work, detailed in Physics 19, 100, uses a network that can be tuned between fully classical and fully quantum behaviour, which let the team study how each hardware architecture shapes performance. The approach treats a neural network as a system of interacting binary variables, and it raises the possibility that these networks could double as diagnostic tools for the quantum processors that run them.

Tunable Quantum Neural Networks with Trapped Ions and Qubits

The team trained the network classically on the MNIST handwritten-digit dataset, a standard machine-learning benchmark, then used quantum hardware to classify images it had not seen before. As they increased the tuning parameter, classification accuracy first improved and peaked in an intermediate regime, before growing more random at higher settings, matching what earlier simulations had predicted. The real hardware, however, added a second source of randomness that those simulations never captured.

That randomness was not always harmful. The measured performance sometimes beat the noiseless simulations, which suggests that a moderate amount of physical noise can occasionally help the network reach the right answer.

Why Noise Shapes the Result on Real Hardware

To understand this, the team examined images the classical network misclassified but the quantum version got right. In some cases the real hardware produced the correct answer even in the classical limit, a setting where an ideal circuit should have failed. Lakhdar-Hamina and colleagues interpreted the effect through an energy-landscape picture, in which ambiguous images sit near competing attractors and a small perturbation can nudge the network toward the correct outcome.

The group also inserted pairs of gates that cancel out in an ideal circuit, and found the effect differed between the trapped-ion and superconducting processors. That confirms a quantum neural network behaves differently depending on the hardware it runs on. It also reframes noise as part of how the network operates rather than a pure obstacle, echoing the controlled randomness used for regularization in classical machine learning.

The experiment was carried out by researchers at the Jülich Supercomputing Centre and the University of Cologne, who are probing the practical limits of quantum neural networks against established benchmarks. As quantum devices mature, tunable networks like this one could become a standard way to measure how real processors affect emerging quantum learning models.