In a new preprint, IQM researchers tested whether a quantum model running on today’s noisy hardware can outperform a full set of classical machine learning models on a classification task designed to favor quantum. They ran the experiments on IQM Emerald (54 qubits) and IQM Garnet (20 qubits), using up to 45 qubits. The quantum model came out ahead in most of the settings tested, and the paper reports where the result stops holding.

 

The experiment

The data is a set of images of T-shirts and dresses from Fashion-MNIST, a standard machine learning dataset. Each image is reduced to a short list of numbers, which are loaded into a quantum circuit running on a chain of qubits. Measuring the qubits at the end gives a set of features for each image. These features are used to calculate how similar any two images are, and a standard classical algorithm, a support vector machine, learns to classify the images from those similarity scores. This is known as a quantum kernel method.

The experiment uses two separate quantum runs, which the authors call the referee and the challenger.

1. The referee runs the circuits on one IQM processor and gives every image a new label. The labels are chosen so that the referee’s quantum kernels predict them well and classical similarity measures predict them poorly, using a method published by Google researchers (Huang et al., 2021).
2. The challenger runs the same circuits independently, on a different processor or a different set of qubits. it gets 100 labelled images for training and predicts the labels of 50 test points.
3. Eleven classical machine learning models get the same 100 labelled images and the same prediction task. They include support vector machines with classical kernels, gradient boosting, random forests and a neural network.

The referee and challenger run on separate hardware so the challenger cannot succeed by reproducing the noise of the device that created the labels.

 

Results at 15, 30 and 45 qubits

At 15 qubits with shallow circuits, the quantum challenger labelled 81% of test images correctly. The best classical model reached 51%, about the rate of random guessing. The 30-point gap was statistically significant under a test designed to limit false positives.

Deeper circuits contain more operations, and each operation adds noise. At the deepest setting tested, the challenger’s accuracy dropped to 68% and the best classical model rose to 64%. The remaining 4-point gap was not statistically significant.

At 30 and 45 qubits, the gap stayed between 17 and 28 points at every depth. For these sizes, the referee and challenger ran on two different qubit chains on the same Emerald chip, chosen to share as few qubit connections as possible.

 

Error mitigation on the deepest 15-qubit circuits

The team tested three techniques for reducing the effect of noise in the 15-qubit case where the gap had disappeared: readout error mitigation, zero-noise extrapolation (ZNE), and adaptive Pauli twirling (APT), a new error suppression technique developed at IQM that will be described in a separate paper.

Readout error mitigation had almost no effect. ZNE and APT both restored a statistically significant gap, of 14 and 17 points respectively.

The team also compared each result with a noise-free simulation of the same circuits, to see whether each technique brought the quantum model closer to its ideal version. most of the improvement in the gap from ZNE came from the referee rather than its closeness to the ideal version: ZNE changed the labels it produced and made the task harder for the classical models, while the challenger’s own predictions improved very little. APT roughly halved the effective error in the quantum features. On labels generated by the noise-free simulation, APT brought Emerald’s accuracy to within 2.5 points of the ideal model.

This comparison shows that an error mitigation technique can increase the measured gap without making the quantum computation better. A noise-free reference makes it possible to check which of the two is happening.

 

Comparison with classical simulation

The authors state that this is not a demonstration of quantum advantage, because the circuits can still be simulated on classical computers. For the largest experiment, 45 qubits at the deepest setting, they simulated the circuits using matrix product states, a standard method for one-dimensional quantum systems. Each circuit took less than 20 seconds to simulate, and the resulting model reached 72% accuracy, compared with 73% for the quantum challenger.

At this scale, direct simulation of the quantum circuits is a much stronger classical competitor than standard machine learning models.

 

Next steps

The authors plan to use circuits based on quantum dynamics that are harder to simulate classically and identify the time evolution of quantum systems as the most promising candidate. The referee and challenger setup, the cross-device checks, and the error mitigation analysis can be reused for those experiments.

The paper also extends the method, so the referee can label new images that were not part of the original dataset. On these labels, the quantum model still outperforms the classical models at most circuit depths, by a smaller margin.

The full paper, Engineered Learning Problems on Noisy Quantum Processors (arXiv:2610.41848), is by Esteban Payares, Nadia Milazzo, Ghada Ben Slama and Thomas Cope. Ghada Ben Slama, now at QbitSoft, contributed during an internship at IQM.