By Dominik Freinberger, MSc
Machine learning has become a key tool for research and industry. However, with growing data volumes and increasing model complexity, conventional computers are increasingly reaching their limits. New approaches could help to overcome these challenges in the future. Quantum Machine Learning (QML) – the combination of quantum computing and machine learning – is seen as a promising solution. QML could set new standards in the future, particularly in data-intensive areas such as medicine and industry.
Contents
Quantum computers use fundamental principles of quantum mechanics such as superposition and entanglement to perform certain computing operations exponentially faster than classical computers. This potential makes them particularly attractive for machine learning, which often requires enormous computing resources. This is precisely where quantum machine learning comes in: It investigates whether and how quantum algorithms could help with complex learning tasks [1].
Under certain conditions, QML algorithms promise more efficient calculations, potentially improved generalization properties or even completely new learning approaches. Possible fields of application range from medical diagnostics to industrial process optimization and financial applications. However, many approaches are still at the experimental stage. The main challenges lie in the susceptibility to errors of current quantum hardware, the limited number of qubits and the open question of the conditions under which quantum approaches to ML problems actually offer advantages.
In order to develop the full potential of QML, application-oriented research, targeted know-how development and strategic investments are needed – also to strengthen technological sovereignty and keep the innovation location competitive in the long term.
A promising approach in quantum machine learning is quantum neural networks (QNNs) – parameterized quantum circuits whose parameters are adjusted using classical optimization methods, similar to classical neural networks. Figure 1 outlines a typical QNN: First, classical input data is encoded into a quantum state of several qubits using date encoding. This is done using special quantum operations, so-called quantum gates, which map the data into an often high-dimensional space. This is followed by a parameterized quantum circuit (variational quantum circuit) that contains further quantum gates with free, trainable parameters. These parameters are adjusted using a classical optimization algorithm in order to minimize a cost function. Finally, a quantum mechanical measurement takes place; this is necessary in order to be able to read out classical information from the quantum model. As with classical ML, the output is compared with a known target value in order to update the QNN parameters.
The advantage of this hybrid quantum-classical architecture lies in the distribution of the computing load: the complex state manipulations are performed on quantum hardware, while the training is carried out using proven classical optimization methods. This means that the first adaptive models can already be implemented on today’s quantum computers, which are still limited and susceptible to interference.
Figure 1: Schematic of a Quantum Neural Network with feature map (reading of classical data), Variational Quantum Circuit (the adaptive component) as well as classical measurement and optimization.
As part of the FFG research project QML4Med [2], we were able to systematically investigate the potential of quantum machine learning in medical data analysis. In addition to the development of methodological know-how, the focus was on comprehensive potential analyses – for example in the application to tabular patient data, for ECG diagnostics or for pathology detection from image data.
A core aspect was the investigation of the resilience of Quantum Neural Network architectures under the influence of noise as it is present on real quantum hardware. In a comprehensive empirical study [3], popular QNN architectures were evaluated in terms of their performance under realistic hardware conditions. Based on this, a new QNN architecture was developed [4], which has a significantly higher robustness against noise due to the exclusive use of natively available quantum gates. This work was presented in a talk and a poster at the international quantum conference IEEE QCE 2024 in Montréal.
Another scientific paper was submitted for the IEEE QCE 2025 conference, which investigated the role of the quantum part in hybrid quantum-classical models as part of a broad benchmark study. The results showed that hybrid models do not automatically perform better than their classical counterparts. Rather, their success depends largely on a careful tuning of the architecture, which requires a sound understanding of both classical and quantum machine learning. The study thus provided valuable guidance for future developments in the field of application-oriented hybrid models.
Quantum machine learning is still in its infancy – but the first application-oriented studies such as QML4Med show the potential that lies in the combination of quantum mechanics and AI. This makes it all the more important to build up expertise at an early stage and identify specific fields of application.
Figure 2: Workflow in the QML4Med project [2] – QML models were comprehensively analyzed in terms of model accuracy and explainability using common medical data types.
[1] J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and S. Lloyd, “Quantum machine learning,” Nature, vol. 549, no. 7671, pp. 195-202, Sep. 2017, doi: 10.1038/nature23474.
[2] “QML4Med: Quantum Computing meets Machine Learning in Medicine”, RISC Software GmbH. Available at: https://www.risc-software.at/referenzprojekte/qml4med/
[3] P. Moser, A. Maletzky, and M. Giretzlehner, “An Empirical Analysis of Realistic Noise in Quantum Neural Networks for Medical Classifications of Tabular, Signal and Imaging Data”, in IEEE International Conference on Quantum Computing and Engineering (QCE), 2024, doi: 10.1109/QCE60285.2024.00191
[4] P. Moser, A. Maletzky, and M. Giretzlehner, “HN-PQE: Hardware-Native Parameterized Quantum Embedding for Noise-Resilient Classifications of Medical Signals and Images,” in IEEE International Conference on Quantum Computing and Engineering (QCE), 2024, doi: 10.1109/QCE60285.2024.10372
Researcher & Developer
