Quantum Neural Networks Face the Hardware Test
Artificial neural networks have become powerful tools for finding patterns in complex data, from classifying images to predicting protein structures and assisting mathematical discovery. Yet their success has so far relied almost entirely on classical hardware. Recent developments in quantum-computing technologies make it timely to ask whether trainable models can also make use of quantum effects such as superposition and the intrinsic uncertainty associated with quantum measurements. What’s more, running neural networks on real quantum processors could potentially turn these networks into probes, revealing how different hardware architectures shape networks’ behaviors. Djamil Lakhdar-Hamina and colleagues at the University of Maryland, College Park, and Duke University in North Carolina have now taken an important step toward characterizing such hardware-dependent effects by implementing a tunable quantum neural network across two leading quantum-computing platforms: trapped ions and superconducting quantum bits (qubits) [1].
A few years ago, the first experimental implementations of quantum neural networks began to test whether theoretical ideas from quantum machine learning can survive contact with real devices [2, 3]. Partly because of the challenge of designing hardware-friendly architectures that can make meaningful use of current quantum devices, these demonstrations remain rare, especially for benchmark tasks such as image classification [1]. They are nevertheless important. Theoretical advantages, such as access to high-dimensional quantum feature spaces [4] or better generalization from few training data [5], do not automatically translate into practical performance on noisy quantum hardware. Experimental benchmarks are therefore needed to determine not only whether a quantum neural network can work, but also which part of its behavior comes from the model and which part comes from the processor.
One route to such a benchmark originates in an older physics view of neural networks. In the 1980s, physicist John Hopfield showed that neural networks can be understood as systems of interacting binary variables, with computation emerging from collective dynamics and attractor-like states [6]. In this language, binary neurons resemble interacting spins, and learning shapes an effective landscape in which different outputs correspond to different stable spin configurations. These properties make binarized neural networks a natural starting point for quantization. In such a network, each neuron outputs only one of two values—much like a spin pointing up or down—which determine the values in the next layer of the network. A quantum version can replace these outputs with qubit measurement outcomes. The outcomes from one layer then determine qubit rotations in the next layer—implemented via sequences of quantum logic gates—whose subsequent measurements produce the outputs passed forward through the network.
Lakhdar-Hamina and colleagues used this idea to construct a quantum neural network that can be tuned continuously between a deterministic classical binary network and a quantum regime (Fig. 1). In the classical limit, the network gives the same answer each time. Away from that limit, quantum effects give the classification process a probabilistic character, so repeated runs of the same input can lead to different classifications. This architecture allowed the team to ask how much quantum measurement uncertainty helps, at what point noise becomes harmful, and how the same quantum-neural-network architecture behaves across trapped-ion and superconducting hardware.
To test the network, Lakhdar-Hamina and colleagues trained it classically on the MNIST handwritten-digit dataset [7], a standard benchmark in machine learning. They then used quantum hardware to classify test images that the model had not seen during training. The researchers increased the tuning parameter that controls the network’s “quantumness.” As they did so, the classification accuracy first improved, attaining its best performance in an intermediate regime, before larger values made the outputs increasingly random. This behavior had been anticipated in earlier simulations of the same architecture [8]. The new experimental element is that the hardware, being real, added a second source of randomness not captured by those simulations. In some regimes, the measured performance lies slightly above the noiseless simulation, suggesting that moderate physical noise can sometimes assist rather than simply degrade the classification process.
The researchers examined this effect more closely using images that the deterministic classical network classified incorrectly but that the quantum version could classify correctly. For such images, real hardware could sometimes produce the correct answer even in the classical limit, where an ideal noiseless circuit would always return the wrong one. Lakhdar-Hamina and colleagues interpreted this behavior using an energy-landscape picture. An ambiguous image may sit near two competing attractors, one associated with the wrong label and one with the correct label. A small perturbation can therefore occasionally push the network toward the better outcome. To test how hardware noise enters this picture, they inserted pairs of gates that should cancel in an ideal circuit. The added noise affected trapped-ion and superconducting processors differently, showing that the behavior of a quantum neural network can depend on the hardware architecture on which it is implemented.
The result highlights a subtle point. In quantum neural networks, noise is not simply an obstacle, but also part of the network’s behavior on real hardware. This insight adds to a broader picture in which moderate noise can sometimes be useful. Classical machine learning already exploits this idea, using noise as a form of regularization [9], and recent theory in quantum machine learning suggests that stochastic noise can also help variational quantum algorithms [10]. What Lakhdar-Hamina and colleagues show, however, is that the effect is not universal. In some regimes device noise can further assist classification, while in others it destroys the useful structure of the computation. Crucially, the outcome depends on the quantum hardware itself.
As quantum devices become larger and better controlled, such tunable quantum neural networks could help clarify how to build more complex quantum learning models. They offer a way to study when measurement uncertainty and hardware noise help or hinder learning and how these effects change from one hardware architecture to another. Such benchmarks will be essential for determining whether quantum neural networks can become reliable tools on real processors rather than just promising models on paper.
References
- D. Lakhdar-Hamina et al., “Benchmarking a tunable quantum neural network on trapped-ion and superconducting hardware,” Phys. Rev. Lett. 137, 040601 (2026).
- A. Abbas et al., “Power of quantum neural networks,” Nat. Comput. Sci. 1, 403 (2021).
- H.-L. Huang et al., “Experimental quantum generative adversarial networks for image generation,” Phys. Rev. Applied 16, 024051 (2021).
- M. Schuld and N. Killoran, “Quantum machine learning in feature Hilbert spaces,” Phys. Rev. Lett. 122, 040504 (2019).
- M. C. Caro et al., “Generalization in quantum machine learning from few training data,” Nat. Commun. 13, 4919 (2022).
- J. J. Hopfield, “Neural networks and physical systems with emergent collective computational abilities,” Proc. Natl. Acad. Sci. U.S.A. 79, 2554 (1982).
- Y. LeCun et al., “Gradient-based learning applied to document recognition,” Proc. IEEE 86, 2278 (1998).
- R. Barney et al., “Natural quantization of neural networks,” arXiv:2503.15482.
- C. M. Bishop, “Training with noise is equivalent to Tikhonov regularization,” Neural Comput. 7, 108 (1995).
- J. Liu et al., “Stochastic noise can be helpful for variational quantum algorithms,” Phys. Rev. A 111, 052441 (2025).