First quantum neural network implemented on two computing platforms

The researchers implemented quantum neural networks on two different quantum computing platforms: trapped ions and superconducting qubits. This is a step toward testing whether such a system can deliver on its theoretical promise. The research, detailed in Physics 19, 100, uses a network that can tune between fully classical and fully quantum behavior, allowing the team to study how each hardware architecture shapes performance. This approach treats neural networks as systems of interacting binary variables, raising the possibility that these networks can also serve as diagnostic tools for the quantum processors that run them.

Tunable quantum neural network with trapped ions and qubits

The team classically trained the network on the MNIST handwritten digits dataset, a standard machine learning benchmark, and then used quantum hardware to classify never-before-seen images. With increasing tuning parameters, the classification accuracy first improved, peaking in the intermediate region, and then became more random at higher settings, consistent with what was predicted by previous simulations. However, the real hardware added a second source of randomness that the simulation could never capture.

That randomness wasn’t necessarily harmful. Measured performance may outperform noise-free simulations. This suggests that a moderate amount of physical noise may help the network arrive at the correct answer.

Why noise affects results in real hardware

To understand this, the team looked at images that were misclassified by the classical network, but correctly classified by the quantum version. In some cases, the real hardware produced correct answers even in classical limits, i.e. settings where an ideal circuit would fail. Lakhdar-Hamina et al. interpreted this effect through energy landscape photography. In this picture, the ambiguous image is located near a competing attractor, and small perturbations can push the network toward the correct outcome.

The group also inserted pairs of canceling gates in the ideal circuit and found that the effects differed between the trapped ions and the superconducting processor. This confirms that quantum neural networks behave differently depending on the hardware on which they are run. It also reconstructs noise as part of the network’s behavior rather than a pure obstacle, mirroring the controlled randomness used in classic machine learning regularization.

The experiment was carried out by researchers from the Jülich Supercomputing Center and the University of Cologne and investigates the practical limits of quantum neural networks by comparing them to established benchmarks. As quantum devices mature, such tunable networks could become a standard way to measure how real-world processors impact emerging quantum learning models.