Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesQuantum machine learning (QML) uses quantum processing as one part of a workflow that still depends on classical computing. A 2026 paper reports classification and time-series experiments using a neutral-atom quantum reservoir, with classical computers handling preprocessing and model training. That is a significant research demonstration—not evidence that quantum computers generally outperform classical machine learning on business or scientific data.
What quantum machine learning means in practice
QML covers approaches that bring quantum computing into machine-learning workflows. In the experiment described in “Large-scale quantum reservoir learning with an analog quantum computer”, the quantum processor did not replace a conventional machine-learning stack. Classical preprocessing prepared the inputs, the quantum system transformed them, and classical models used the resulting measurements to make predictions.
This division of labor matters. The useful question is not simply whether a project “uses quantum,” but which operations ran on quantum hardware, what remained classical, and whether the complete workflow improved on a suitable classical alternative.
How the neutral-atom reservoir workflow works
- Prepare the input classically. The data is encoded for the quantum system. Depending on the task, preprocessing may include feature engineering or dimensionality reduction.
- Process the input with the quantum reservoir. The neutral-atom analog system evolves, and its state is probed through repeated measurements. The method uses the system’s responses as features rather than repeatedly tuning quantum-circuit parameters through a hardware optimization loop.
- Train and predict classically. The measurements are converted into embeddings—numerical representations that a classical model can use. The paper describes classical training approaches including linear support vector machines and regression.
The approach is called gradient-free because it avoids the parameter-optimization loops used by many variational quantum methods. That can sidestep some costly quantum gradient-estimation work; it does not remove the need for data preparation, measurements, classical computation, or careful evaluation.
#1 Best Overall
What the experiment reports—and what it does not
The paper reports classification and time-series prediction experiments on neutral-atom analog hardware, with effective learning observed at system sizes up to 108 qubits. Its authors describe this as the largest quantum machine-learning experiment to date; that superlative is their characterization of the work reported in the paper, whose arXiv record lists an initial submission on July 2, 2024, and a revised date of August 24, 2026.
One task-specific result was a test accuracy of 0.935 for binary classification of the MNIST digits 3 and 8, using 220 measurement shots. This figure belongs to that particular dataset, task, and measurement setting; it is not a general accuracy estimate for QML.
Rank #2
The paper also reports a comparative quantum-kernel advantage on synthetic datasets constructed around geometric differences between generated quantum and classical data kernels. The result is bounded by that deliberate dataset construction and comparison. It does not establish an advantage on ordinary commercial workloads, nor does it show that QML broadly beats classical machine learning on naturally occurring data.
How to distinguish the laboratory result from industry claims
A separate 2024 EE Times interview by Pablo Valerio features Kristen Gilkes, EY’s Global Innovation Quantum leader, and Marta Estarellas, CEO of Quilimanjaro Quantum Tech. Their examples describe industry views and projects; they are not results of the neutral-atom reservoir experiment.
| Evidence | What is described | How to read it |
|---|---|---|
| Primary research paper | Classification and time-series prediction experiments on neutral-atom analog hardware; up to 108 qubits. | A reported experimental study, including task-specific results and a synthetic-data kernel comparison. |
| Gilkes’s interview examples | Satellite imagery analysis for fire detection, farming, and insurance claims assessment; a garbage-truck optimization project on a small island. | Examples and claims attributed to Gilkes in the 2024 interview, not independently established comparative results in the cited experiment. |
| Estarellas’s interview examples | Supply-chain constraints framed as binary constraint-optimization problems, and the need to direct suitable work to a quantum processing unit. | Industry perspective on possible applications and system design, not proof of general quantum advantage. |
Gilkes described the field as being in “the stage of quantum utility” and said quantum computing “already provides practical value and solves real-world business problems.” That is her position in the interview, not a conclusion independently demonstrated by the narrower paper. Estarellas’s integration point is more concrete: “You need to have a hardware orchestrator that identifies which part of the problem makes sense to send to the QPU [Quantum Processing Unit].”
Why a promising demonstration is not yet broad practical advantage
The paper discusses noise, finite measurement resources, training challenges in contemporary quantum methods, and the costs of gradient estimation. Its reservoir design avoids some optimization burdens, but it remains experimental. A system’s qubit count alone does not establish useful performance: the result also depends on the task, data representation, measurements, preprocessing, and the comparison being made.
Integration is another practical hurdle. A useful hybrid system needs software and infrastructure that can decide which parts of a workload belong on quantum hardware, route those operations, and connect their outputs to classical applications. The interviewees emphasize this orchestration and application-layer work; it is distinct from demonstrating a quantum model on a research task.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to evaluate a QML claim
When assessing a paper, product announcement, or application claim, look for enough detail to answer these questions:
Recommended Free Tools
Best Value
- Task and data: Is the workload classification, forecasting, optimization, or something else? Is the dataset synthetic or observed?
- Hardware: What architecture and system size were used, and under what experimental conditions?
- Quantum contribution: Which computations ran on the quantum processor, and which remained classical?
- Baselines: Were suitable classical methods evaluated on the same task, with their settings tuned fairly?
- Resources: What measurement-shot count, runtime, preprocessing effort, and noise conditions were involved?
- Strength of evidence: Is the result a proof of concept, a task-specific improvement, an advantage on a constructed dataset, or evidence supporting a broader claim?
Those distinctions put the 108-qubit experiment in context: it demonstrates a hybrid learning method at notable research scale, while its synthetic-kernel finding and task-specific benchmarks do not settle whether quantum processing offers a practical advantage across everyday machine-learning workloads.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
