Quantum Classifier Demonstrations for Data Scientists

*How can you use quantum neural networks (QNNs) to gain a quantum advantage on classical data?* We propose to use QNNs (and other quantum algorithms, including quantum signal processing) to process data in quantum sensors. Attempts over the past 7+ years to find near-term practical applications of quantum neural networks on classical data have faced a variety of challenges, including: if the classical data is small enough to be able to load into a quantum computer, then it has (empirically) always been possible to address the same problem with a classical neural network - and without the downsides of quantum computing with current (noisy) hardware. Rather than trying to tackle problems in the setting where the classical data originates from a classical computer's memory, we switch the framing of the problem slightly, but in a way that makes a huge difference: what if we use QNNs to perform classification on classical but a priori _unknown_ data? What do we mean by _unknown_ data? A quantum sensor senses a classical signal that is unknown to us, but is ultimately classical. We can use a QNN to help reveal a _trained nonlinear function_ of the unknown classical signal. One of the examples we have explored shows how you can gain an advantage where both the quantum sensing and quantum computing are performed by a single qubit! If you already knew the classical signal, there would be no hope for a quantum advantage (simulating a single qubit is of course trivial), but in the sensing setting we don't know the signal a priori. We have been able to show it is possible to gain a quantum computational-sensing advantage using quantum signal processing (QSP) treated as a QNN, versus first using a conventional quantum sensor and then postprocessing to compute the nonlinear classification function classically. By performing an approximation of the nonlinear classification function in the quantum system before measurement, the quantum sampling noise is greatly reduced: measurements of the system yield 0 or 1 with high probability depending on which of two classes the signal was in. We have a preprint on the arXiv showing various schemes for quantum computational sensing with a small number of qubits and/or bosonic modes, tested on a variety of binary and multiclass classification problems: https://lnkd.in/enQxFDNt I am optimistic about the prospects for experimental proof-of-concept demonstrations given the modest quantum resources required (down to just a single qubit and a not-particularly-deep circuit). Congratulations to Saeed Khan and Sridhar Prabhu, as well as Logan Wright!

Photonic quantum classifiers with provable generalization guarantees Most quantum machine learning models look impressive on paper but stumble on a basic question: can they generalize from limited data? The culprit often hides in how inputs enter the circuit. In "data reuploading" schemes, classical features are repeatedly encoded into a qubit and interleaved with trainable rotations—an elegant idea that turns a single qubit into a universal classifier. But most experimental implementations merge encoding and training into one gate, quietly inflating model complexity until generalization guarantees vanish. Martin Mauser and coauthors revisit this design choice on a femtosecond-laser-written photonic processor, implementing the original proposal with encoding and trainable rotations kept strictly separate. Each layer is a pair of Mach-Zehnder interferometers acting on a single-photon qubit in dual-rail encoding, with output probabilities fed into a linear discriminant analysis. The ML contribution is sharp. They prove that the compressed variant used in most prior experiments has infinite VC dimension—no finite training set guarantees generalization—while their separated architecture has VC dimension 2L+1 for L layers. That's the difference between a model that memorizes anything and a PAC-learnable classifier with provable error bounds. The loss landscape tells the same story: Hessian sharpness ~0.23 for the separated scheme versus ~7×10¹² for the compressed one. Flatter minima, easier training, better generalization. Experimentally, the photonic classifier handles circles, moons, tetromino letters, and Overhead MNIST (ships vs. cars, PCA-reduced to 20 features), beating plain LDA baselines and staying robust to photon-counting noise. The same circuit also runs on coherent light, opening a path to low-energy optical inference. For applied R&D teams, the lesson cuts across quantum and classical ML: architectural choices that look like minor implementation details can silently destroy a model's ability to generalize. In drug discovery, materials screening, and biotech pipelines—where labeled data is scarce—favoring models with provable VC bounds and smoother loss landscapes isn't theoretical hygiene; it's what determines whether a prototype survives contact with new compounds, alloys, or assays. Paper: Mauser et al., Science Advances (2026) — CC BY 4.0 | https://lnkd.in/eUEqucer #MachineLearning #QuantumMachineLearning #QuantumComputing #PhotonicComputing #AIforScience #DataReuploading #DeepLearning #GenerativeAI #VCDimension #Generalization #DrugDiscovery #MaterialsScience #Biotech #ArtificialIntelligence

🔮 Can quantum computing improve how we assess risk in insurance? This week I built a compact demo exploring that very question — using quantum kernels to classify insurance policyholders based on real-world features like: 🚗 Vehicle age 📅 Customer tenure 🛣️ Annual mileage These variables form a 3D feature space that we encode into a quantum state. The quantum kernel then measures the fidelity (overlap) between states, allowing an SVM classifier to learn patterns that classical kernels might miss — even in small datasets. The results? 👉 A clean separation between “likely claim” (red) and “no claim” (blue) in the 3D risk landscape — powered entirely by quantum feature maps. This approach shows how quantum-enhanced kernels could augment traditional actuarial models, especially in domains with complex nonlinear risk factors, limited samples, or strong feature entanglement. Quantum insurance modeling might sound futuristic — that’s why I built this prototype: to show how Quantum Machine Learning (QML) can be interpretable, transparent, and domain-specific. #smellslikequantumspirit #avenue78

Published today in Nature Portfolio’s npj quantum information journal: Quantum Machine Learning using 156 qubits🎉 Very happy about this one. Last year, our E.ON Quantum team in collaboration with IBM Quantum successfully demonstrated multi-class quantum classification for the Vehicle-2-Grid problem where we predicted charge and discharge amounts for electric vehicles to reach a certain total energy volume that could be sold in advance in the day-ahead energy market. Along the way, we had to develop a novel #errormitigation technique specifically designed for the underlying quantum ML method used here called quantum kernels. This allowed us to run the #largest experimental implementation of quantum classification, or more generally, any quantum machine learning algorithm on IBM quantum computers (arguably in the world). I really like this project because it showcases the innovation that can occur when starting off with a well-defined #realworld business problem, and the resulting science to make it happen can be published in high impact journals such as this. Thanks for the great collaboration! 🎉 Gabriele Agliardi Giorgio Cortiana Anton Dekusar Kumar Ghosh Naeimeh Mohseni Víctor Valls Kavitha Yogaraj Sergiy Zhuk Nico Einsidler Raja Hebbar Scott Crowder Sarah Sheldon Jay Gambetta E.ON Digital Technology (Article link in the top comment)